Changelog

Track the latest updates and improvements to HyperDevs.

v0.1.34

Latest Release November 2026

Added

  • • 3D Edit’s AI builder (beta). 3D Edit now has a chat that builds props for you. Describe one, such as “a wooden treasure chest”, and the AI you use builds it from simple parts, then opens it in 3D Edit, ready to change by hand. The AI writes the list of parts and HyperDevs builds them itself, so no code runs, and building uses no credits, only the chat. If the AI can see images, it looks at two pictures of each build and fixes what looks wrong, and when you ask for changes, it builds a new version. A bar under the view keeps every version, with your own model and edits, and the AI’s versions come back when you restart the app: Save to library keeps the one you like, and Open in Optimize takes it on to your game. The empty chat has examples to start from. For now the builder makes new models only, and simple props work best.
  • • The Store. A new tab, right after 3D Edit, of free 3D models for your game. Browse it without signing in; to get a model, sign in and click Get. It goes into your library, with no credits used, and its page then offers Send to Game, Open in 3D Edit and Show in Viewer. A model you already have says In your library. For now only the HyperDevs team adds models, and more come over time. With ten tabs, the tab bar switches to icons a little sooner as the window narrows.

Improved

  • • No step limit for Hyper 30. A Hyper 30 chat that builds in Studio is no longer stopped at 30 steps: like Claude’s, a long build keeps going until it’s done, within the 3 hours a Hyper 30 reply can now run, up from an hour. The Image Gen, 3D, Animation and 3D Edit chats keep their limit of 15 steps on every model.
  • • Two Hyper 30 chats at once. Now that Hyper 30 runs in the cloud, two of your chats can get replies at the same time. Before, the second still waited for the first to finish, saying “Your other chat is using Hyper 30”.

Fixed

  • • Replies through a sleep. A reply still running when your PC went to sleep, such as a long Hyper 30 build, ended as soon as the PC woke, as if it had stalled: the time asleep counted against the reply’s limits. It no longer does, so the reply carries on after the PC wakes.
  • • The AI sees Studio’s errors. When code the AI ran in Studio failed, the AI was told only “Code execution failed”, without Studio’s error, so it couldn’t tell what to fix. It now gets Studio’s own error message and can fix the code.
  • • Studio calls that never come back. With Studio connected but not answering, a call the AI made to it could hold the reply up for as long as 25 minutes. A call Studio doesn’t answer now ends after a few minutes, and the AI is told to ask you to check Studio.
  • • Picking colours in 3D Edit. 3D Edit has its own colour picker now, with an eyedropper that picks up the model’s real colour. The old picker’s eyedropper read the colour off the screen, where the model is lit and shaded, so each pick came out a little different and colours drifted as you painted with them.
  • • Turning the view in 3D Edit. Right-drag, and dragging with the View tool, turned the view around the point you pressed on, so going round a model in a few drags walked it off to one side. It now turns around the centre of the view, as the app’s other 3D views do.
  • • Painting after a cut. Painting the flat face a cut fills in did nothing. Each such face now gets room of its own on the model’s texture, so it takes paint from the first stroke.
  • • The meter when free use runs out. When today’s free Hyper 30 use was used up, the meter under the message box showed a full grey bar. It now turns red, as does the credits meter once you reach your daily cap.
  • • A HyperDevs folder behind a link. If your HyperDevs folder sits behind a link, or in a folder Windows moved, the app couldn’t open the models and pictures saved in it, and the Store couldn’t find your library’s models. It can now.

v0.1.33

November 2026

Added

  • • Hyper 30 is moving to the cloud. Replies get faster, with no waiting in line, and everyone gets free use every day, with the option to keep going with credits after it.

Improved

  • • 3D Edit is faster on big models. Turning the view should be quicker on a closed, double-sided model, such as some Blender exports, as 3D Edit now draws only its outside; thin parts like leaves or cloth still show from behind. While you paint, the view shows a lighter copy of the model until you let go, so strokes should be smoother; the shape looks a little coarser only while the button is down. And textures larger than 1024 × 1024, the most Roblox shows, now open fitted within it, as Optimize already makes them, so a model you save from 3D Edit has them at that size.
  • • Send to Game shows each step. While Studio builds a model you sent, the 3D tab’s bar and the Viewer’s card now move through the build’s steps, such as “Adding vertices” and “Creating MeshPart”, as Optimize does. Before, they said “Sending to Studio…” until it was done.
  • • Lighter panel drags in the Game tab. Dragging a divider in the Game tab, or resizing the window, redrew the whole tab, chat, Hierarchy and Properties included, for every pixel you moved. It now redraws only when the tab’s header has to change its layout, a few times in a long drag, so dragging should feel smoother.

Fixed

  • • The AI waits for its tools. The AI stopped waiting for any of its tools after 20 seconds, so 0.1.32’s 25-minute wait for generations from the chat never took effect. A 3D model, animation or image the AI started went on, finished and used your credits, while the AI was told it had failed and might try again. Anything it did in Studio that took longer failed, and a permission prompt or a script change waiting for your OK closed by itself. The AI now really waits: up to 25 minutes for a generation, the full 2 minutes for code it runs in Studio, and up to 5 minutes for your answer to a permission prompt or a script change.
  • • Script changes call you back. When a script change from the AI is waiting for your Apply while HyperDevs is minimized or behind another window, the taskbar now flashes and a notification shows, as for a permission prompt. A script change nobody answers within 5 minutes closes without writing anything, and the chat tells you, so you can ask again. And when a prompt or a script change goes unanswered, the AI’s requests queued behind it for the same thing end with it, instead of each coming up in turn.
  • • Keys under a permission prompt. With a permission prompt over a script change, one Escape refused the prompt and also rejected the script change under it, and Ctrl+S applied that change. Escape now answers only the prompt, and a script change’s shortcuts, Ctrl+S included, work only when nothing else, such as a bug report, is open over it.
  • • The permission prompt over open dialogs. A permission prompt that came up while Settings or a bug report was open was drawn behind it, yet took your clicks and keys, so Settings looked frozen and Escape refused a request you couldn’t see. The prompt now comes up in front, and a bug report you were writing is still there once you’ve answered it.
  • • Two places open in Studio. With two places open, the AI could make part of a change in the place it named and the rest in the one shown in the Game tab, and read back from the wrong one. Everything it does now goes to the place it names, and when the reply ends, the playtests it left running are stopped in every place, not just one.
  • • The Game tab keeps up with Studio. When a playtest ended in Studio itself, for example with Shift+F5, the Game tab still showed it playing, and its Stop said the playtest did not stop. The tab now checks with Studio when you come back to HyperDevs, when a Game chat reply ends and when Studio connects. After a build, Properties now show the new values, and what you selected in Studio during the build stays selected instead of snapping back.
  • • The AI engine after a failed start. If the AI engine didn’t start the first time when you opened HyperDevs, a later try could bring the AI back, with its status light green, but without any of its tools for the rest of the session. The tools now come up with whichever try works. An engine that quits while starting is now started again by itself, and a failed start no longer leaves an extra copy of the engine running, even after you quit.
  • • Models with KTX2 textures. A model whose textures are KTX2 (Basis) images showed without them in the Viewer, 3D, Optimize, Animation and 3D Edit tabs. Its textures now show. 3D Edit shows them, but can’t paint on them.
  • • Studio’s beta question, one at a time. When a queued Send to Game started while Upload to Roblox was asking about Studio’s “CreateAssetAsync Luau API” beta, the send’s question took the upload’s place: the upload ended without a word, and its Never mind cancelled the send instead. Each question now waits its turn and opens as a popup of its own.
  • • Switching accounts mid-task. Signing out, or another account signing in, while an image or model was still saving could carry it over to the next account: messages about it showed there, Make 3D handed the image on to the 3D tab, and the 3D tab’s Send to Game or Send to Prep went on to send the model to Studio or open it in Optimize. Studio’s beta question from the last account’s Send to Game could also stay on screen, and Send anyway built that model. None of this carries over now; the file itself still lands in your project.
  • • Report a bug with a dialog open. With Settings or another dialog open, clicking Report a bug closed the dialog without opening the report, and Minimize, Maximize and Close did the same. The report now opens on top, with the dialog still there when you close it, and Minimize, Maximize and Close just work. The credits, account and sign-in buttons now close the dialog in front and open their own, where they opened nothing, and a script change shown at the side no longer covers the buttons at the top right.
  • • The tour’s close button. At some window sizes the tour’s card sat over the strip you drag the window by, so its close button did nothing and pressing the card moved the window. The card now stays below the tab bar.
  • • Room for nine tabs. The tab bar was still sized for eight tabs, so since 3D Edit made nine, some window widths left no room to drag the window by the tab bar, and the buttons at its right end could run past the window’s edge. It now switches to shorter labels, then to icons, at the widths nine tabs need, so you’ll see them a little sooner as the window narrows.
  • • Generations that are slow to send. A generation whose request took more than 30 seconds to send, such as a 3D model made from a large image on a slow connection, used your credits and never showed up. HyperDevs now keeps track of it while it sends and shows it when it’s done.
  • • Smaller fixes. A Hyper 30 message sent just as HyperDevs renewed your Hyper 30 access in the background no longer stops at once with an empty reply; it waits a moment and goes out. And the Memory scan’s messages now suggest another provider by the same rule as Hyper 30’s, so the two no longer disagree about which you can switch to. Deleting your account now also closes any credits checkout you left open.

Security and privacy

  • • Further security improvements. We don’t publish the details.

v0.1.32

November 2026

Added

  • • The Image Gen, 3D and Animation chats make what you ask for. Ask the chat in one of these tabs for an image, a 3D model or an animated character, and it now makes it instead of only giving advice. The 3D and Animation chats make the picture to start from first when they need one. Each result lands in its tab and your assets even if you switch tabs while it works, and uses credits as the tab’s Generate button does. The Image Gen chat uses the model picked in the tab; to use another, pick it in the tab’s Model menu, which shows its price.
  • • 3D Edit (beta). A new tab, right after 3D Models, for fixing AI-made models by hand: smooth out lumps and bumps, pull parts into a new shape, paint or recolour them, cut a model into pieces, and clear away junk and floating bits. It runs on your PC and costs no credits, and saving makes a new model, so your original stays as it was. It’s a beta: big models can feel slow, so save often.
  • • Hyper 30 can see. Hyper 30 now looks at the Studio screenshots it takes while it builds, so it can check its work the way you would: a part floating or sunk into the ground, a gap, or a menu that never shows up for players.
  • • Queue models for Send to Game. Send to Game no longer turns a model away while another is on its way to Studio. The new one waits its turn and shows “Queued” where you sent it, with Cancel. Models from the 3D tab, the Viewer and Optimize go one at a time, in the order you sent them. If you’re signed out, or another account signs in, while models are waiting, a message says how many weren’t sent, along with any generation that was stopped.
  • • Buy credits from 13. Buying credits now opens at 13, the same age as an account, where it was 18. If you’re under 18, you tick a box to say a parent or guardian agreed before you buy.
  • • An API token for LM Studio. If LM Studio’s server requires an API token (its Require Authentication setting), add the token in Settings → AI Providers → LM Studio. HyperDevs sends it, and says when it’s missing or wrong, instead of “LM Studio didn’t answer”.

Improved

  • • Hyper 30 stops when the job is done. Ask Hyper 30 how to do something and it now answers without changing your game. On a build, it checks its work once and stops when what you asked for is there, instead of running on to its step limit. It still playtests any script it writes.
  • • Hyper 30 thinks in shorter steps. Each step of a Hyper 30 reply now thinks for at most about seven minutes before it acts, where one could think for over a quarter of an hour, and Hyper 30 now uses the settings its model’s makers recommend. A build can also take 30 steps instead of 15 before it stops and asks you to continue.
  • • Faster, lighter Send to Game. A model now goes to Studio in far fewer, fuller steps: a typical 10,000-triangle model takes about 40 instead of about 90, so a send should take roughly half as long. It also no longer passes through the app’s window as a copy, about 20 MB less per send for such a model, so sends use less memory and the app stays responsive.
  • • Optimize in three clicks. A model you send to Optimize now opens ready to optimize: pick a preset, then Send to Game or Save right under the result. You can also send or save a model as it is, without optimizing, including one you dropped or browsed in. Texture and Export are optional steps now, and a model too big for one send says so before it’s queued.
  • • Send to Game shows its progress. While a model is on its way to Studio from the 3D tab or the Viewer, you now see what’s happening (reducing, preparing, building), and its button stays busy until it’s done, even if you switch tabs and come back. The 3D tab’s Send to Game, Send to Prep and Clear buttons now also match the app’s other buttons.
  • • Longer Hyper 30 steps. Each step of a Hyper 30 reply can now write up to 40,000 tokens, up from 32,000, so a long think has more room to reach its answer. When a step is cut off at that limit, the chat now says so plainly, instead of saying Hyper 30 stopped while it was still thinking, or ending with no note. If Hyper 30 had already done part of the work, it says to send “continue”.
  • • How the AI checks its work. Models that can’t read pictures were still offered the screenshot tools, so they took screenshots to check their work, learned nothing from them, and moved the camera for more. A model now gets those tools only if it can see images, or if you turn them on in MCP Tools. The AI is also told that its checks inside Studio run with a plugin’s rights, so they can pass where your game’s scripts would fail: it checks game behaviour in a playtest instead, and uses the game’s own switch for Roblox’s backpack and other built-in UI.
  • • No step limit for Claude. With your own Claude key, a chat that builds in Studio is no longer stopped at 15 steps: a long build keeps going until it’s done, within the 35 minutes a reply can run. Other models keep their limit.
  • • 3D previews rest when nothing moves. The 3D, Optimize, Viewer and Animation previews used to redraw constantly, even with nothing moving, which kept your CPU busy whenever a model was on screen. They now redraw only when something changes: you turn the view, a clip plays, or a model loads. Optimize also keeps its preview up through a run instead of reloading the model each time.
  • • Lighter optimized models. Optimize now fits a model’s textures within 1024 × 1024, the most Roblox shows, so the preview uses far less memory and a model with large PNG or JPEG textures comes out at less than half the size on disk. What Send to Game puts in Studio looks the same.
  • • Repeat optimizes finish at once. Optimizing the same model to the same target again now reuses the earlier result instead of running again. HyperDevs also deletes its own optimize files once they’ve gone a week unused, so they no longer pile up.
  • • Send to Game asks every time. While Studio’s “CreateAssetAsync Luau API” beta is off, Send to Game now asks before every send, so you’re reminded each time that the model will be gone once Studio restarts. Before, after you chose Send anyway, it didn’t ask again until Studio restarted.
  • • The Viewer remembers its filter. The Viewer now opens on the filter and sort you used last (All, Images, Models and so on), after a restart too.
  • • A faster start. Settings and the bug report window now load the first time you open them, not at launch.
  • • Loading a local model. While Ollama or LM Studio loads a model, which can take minutes, the chat now says “Loading” and the model’s name instead of “Thinking…”.
  • • Only the tools the AI can use. The MCP Tools window and the chat’s tool count no longer list Studio tools the AI is never given.
  • • Tool calls written out as text. When a local model writes a tool call out as text instead of making it, the AI is now told the call didn’t run, for the formats most local models use.

Fixed

  • • Long generations no longer get lost. A 3D model, animation or other generation that took long enough could be reported as failed while it actually finished, and one still running when you closed HyperDevs, or lost your connection, was never delivered. HyperDevs now keeps track of each generation you paid for: one that finishes after the app stopped waiting is saved to your assets the next time you open HyperDevs on that account. Closing HyperDevs while one is still being made now asks first and tells you this. If one fails, you’re told whether its credits came back.
  • • Replies that stopped short. A reply that reached its step limit could stop mid-sentence, without the note that you can send “continue”, and couldn’t be rated. It now keeps its ending and that note, and you can rate it. And when a model ran its tools and then stopped without writing anything, the chat showed the tool cards and nothing else, as if it hadn’t answered: it now says the model stopped there and that you can send “continue”.
  • • Signing in keeps your message. If you typed a message and were asked to sign in first, signing in no longer clears it, or anything else you had on screen.
  • • Stopping a generation. In 0.1.31, Stop on a generation fal.ai had already started still used your credits but threw the result away. It now finishes and is saved to your assets, and a message says so; a generation that hasn’t started yet is still cancelled and refunded. The chat’s Stop now works the same way on a generation the AI asked for, and after Stop the 3D and Animation tabs no longer say the generation failed.
  • • Reducing a model. Reducing a large AI model, such as some Trellis 2 models, could leave it empty: Optimize said it worked, the model vanished from the preview, and Send to Game failed. The model now keeps its main part and lands close to the triangles you asked for. Reducing could also smear or scramble a texture, on up to a third of the surface for some models; it now keeps the texture’s layout, so far less of the model comes out the wrong colour.
  • • Optimize keeps your work. Leaving the Optimize tab used to empty it, even while an optimize or a Send to Game was still running. It now keeps the model, the result and the step you were on until you load another model, and work that finishes while you’re away shows up when you come back. A send’s progress also shows only with the model it’s for.
  • • Waiting for Hyper 30. Two of your chats could reach Hyper 30 at once and get in each other’s way: chats waiting for it all sent the moment it was free, and a message sent while your other chat was still getting a reply went out alongside it. Your chats now take turns, in the order you sent them, and a chat waiting behind your other one says so. After you cancel a wait, the chat still shows that someone else is using Hyper 30, instead of looking free, and a chat waiting for Hyper 30 keeps its tools if you pick another model meanwhile.
  • • When the AI needs your OK while you’re away. If the AI asks permission to change Studio while HyperDevs is minimized or behind another window, the taskbar now flashes and a notification shows. A request nobody answers within 5 minutes closes, and the chat says the change didn’t run, instead of the build quietly failing.
  • • Switching tabs while Image Gen or Animation works. Leaving Image Gen while an image was generating stopped it, and once fal.ai had started it, the credits were spent with no image. It now keeps going and lands in the tab and your assets, as 3D models already did. The Animation tab now saves a model it makes while you’re on another tab as soon as it’s done; before, it waited until you came back, and never saved it if you closed HyperDevs first. The tab also keeps its list of animations, and Stop still works after a switch.
  • • Tables in replies. A table in the AI’s reply showed up as lines of | marks. It now shows as a table, and a wide one scrolls sideways inside its message. Plain web addresses become links.
  • • Signing in and credits in the generation tabs. Image Gen, 3D and Animation now open the sign-in window when you’re signed out, instead of Settings or an error. Running out of credits shows one message and Buy credits, also part-way through Generate Views or a batch of animations, instead of one error per view. And when the server fails while setting your credits aside for a generation, HyperDevs now says something went wrong and that nothing was charged, instead of saying you didn’t have enough and offering to sell you more.
  • • Buying credits. After a purchase, the confirmation now names the credits you bought, and no longer says they haven’t arrived when they have, for example when you switched back to HyperDevs before the browser asked to open it.
  • • Playtests left running. When a reply ended at its step limit or on an error while a playtest the AI had started was still running, the playtest kept going in Studio. HyperDevs now stops it when the reply ends.
  • • Long generations from the chat. The AI gave up on a generation after 10 minutes, though a 3D model can take up to 20. It now waits up to 25 minutes, and to make room, a reply on any model but Hyper 30 can run 35 minutes instead of 25.
  • • Better advice when Hyper 30 can’t answer. Hyper 30’s messages, and the Memory scan’s, now suggest another provider only when you’ve set one up; otherwise they say when to try again.
  • • Clearer first steps. The example prompts in the 3D, Optimize and Viewer chats are now questions those chats can answer, and the sign-up screen no longer mentions credits.
  • • Add to Game with .gltf models. Add to Game in the Viewer failed on a model saved as .gltf, saying it couldn’t be reduced. It now sends .gltf models as it does .glb ones.
  • • Animation’s walk and run. A rigged character made in the Animation tab now keeps the walk and run animations that come with it, and each extra animation is named for what it is. With the Idle animation picked, Save to Assets and Send to Game no longer stay greyed out. Asking for the same animation twice in one rig is now charged once, as only one is made. And a model that came back without a rig, such as a chair, is no longer labelled Rigged and Animated in the Viewer.
  • • Image settings. Flux 2 Flex images get their full 28 steps again (they had dropped to 4 and came out rough), Flux Pro 1.1 Ultra follows the aspect ratio you pick instead of always making 16:9, and a step count a model can’t take is fitted to its range instead of failing.
  • • Studio’s status during a long build. The Game tab’s Open a game card, its status dot and Settings’ Connected dot no longer disagree with the green status light.
  • • Studio tools in a narrow Game tab. With the Game view narrow and a long place name, the name crowded out most of the header’s Studio tools. The tools now keep their room, and the place name and the view switch’s labels shorten first.
  • • Gemini’s daily limit. When a free Gemini key runs out of its daily requests, the chat now says so after about two minutes, with when they come back, instead of retrying for nearly nine.
  • • Claude 5’s thinking. A long think on Claude 5 could end as “not responding”. Its thinking now shows in the chat as it goes.
  • • Long replies on Ollama and LM Studio. When a long reply outgrew what a local model can take in at once, Ollama quietly dropped the start of the chat. HyperDevs now checks every step, and stops the reply with a note saying what to change.
  • • The model menu after switching providers. After you switched away from Ollama or LM Studio, the next provider’s model menu could be empty, or show the local server’s models, until you restarted HyperDevs. Each provider now shows its own models.
  • • Base URLs in Settings. If you saved OpenRouter’s address and key in the OpenAI row before 0.1.31, HyperDevs now moves them to the OpenRouter row by itself and switches your chats to OpenRouter. And emptying a provider’s Base URL now puts its usual address back.
  • • HyperDevs on OpenRouter. Chats through your OpenRouter key now show up as HyperDevs on your OpenRouter activity page, instead of OpenCode.
  • • A failed script change. If applying the AI’s script change failed, switching the diff’s view lost the error and its Retry button. They now stay.
  • • Bug reports that didn’t send. A report that failed to send no longer counts toward the limit of five an hour.
  • • Smaller fixes. The maximize button on the update screen and the other full-window screens now says Restore when the window is maximized. Export Texture now keeps a texture’s real file type; it named every file .png, so some programs wouldn’t open a WebP or JPEG texture. An error message that already ends in a full stop no longer gets a second one, a failed restart of the AI engine no longer tells you twice to restart the app, and a SurfaceLight shows its own icon in the Hierarchy.

Changed

  • • Hyper3D Rodin 2.5. Rodin 2.5 replaces Rodin v2 in the 3D tab, at the same 60 credits. Multi-view’s Hyper3D Rodin runs on Rodin 2.5 too, still at 65 credits.
  • • Multi-view’s image models. Flux Pro v1 is gone from the list, as fal no longer offers it, and Flux 2 Flex now carries its own name.
  • • Apply Texture to Studio Model. Optimize no longer has this button: it could put the texture on the wrong part of your place. Send to Game already puts the texture on every piece of your model.
  • • The Encryption Key setting. Settings → General no longer has it, and Ctrl+Shift+R no longer does anything. Your saved model choices carry over.

Security and privacy

  • • Updated Terms and Privacy Policy. They now cover buying credits from 13, with a parent’s or guardian’s OK under 18, and say what deleting an account keeps. You’ll be asked to accept them once.
  • • Deleting your account. It now also deletes the bug reports you’ve sent us. The delete screen says what is kept: a record of any request our content policy blocked, without what you typed, and for 120 days a record of each purchase, in case a payment is disputed.
  • • Further security improvements. We don’t publish the details.

v0.1.31

November 2026

Added

  • • Send to Game builds your model in Studio. The 3D tab’s Send to Game and the Viewer’s Add to Game now build the model straight into your open place, reduced to about 10,000 triangles first so Studio can take it. When Studio isn’t connected, the file is ready to drag in as before, and the app says why.
  • • Send to Game needs no Roblox key. Roblox Studio now saves the model and its texture to the Roblox account you use in Studio. If Studio won’t save them, the model still goes in, and the app says it will disappear when Studio restarts and how to fix that. Optimize’s Upload to Roblox goes through Studio too, falling back to a saved Roblox key when Studio refuses or isn’t open.
  • • Send to Game checks Studio first. Keeping a model in your game needs one Studio setting: the “CreateAssetAsync Luau API” beta, under File → Beta Features. When it’s off, HyperDevs says how to turn it on before sending, and you can send anyway: the model goes in, but it’s gone once Studio restarts. Just turned it on? It reminds you to restart Studio first. The green message after a send now shows only for a model that stays.
  • • Ollama and LM Studio on another computer. A Base URL on your home network, or over Tailscale, now works for Ollama and LM Studio. Settings also fixes a typed address with stray spaces or a missing “http://”.
  • • API keys checked when you save them. After you save a key, Settings asks the provider whether it works and says straight away if it was refused. Spaces copied in around a key are removed.
  • • Get Ollama or LM Studio. The provider menu marks Ollama and LM Studio “not installed”, “not running” or “no model”, and the chat links to their downloads.

Improved

  • • Local models get the room they need. Before a message goes to Ollama or LM Studio, HyperDevs checks how much the model can take in at once. LM Studio loads a model that isn’t loaded yet with enough room. When a model has too little, the chat says what to change (for Ollama, its context length), instead of the model quietly losing most of its instructions. Each Ollama model now uses its own recommended settings, and a local model gets up to 8 minutes to start answering.
  • • Your own keys, better used. Gemini and Claude models now get their full context and can see images, and Gemini’s thinking shows in the chat. Each model now uses the settings its maker recommends, and OpenRouter reserves less of your credit for each request.
  • • Clear answers when a provider says no. When a provider says you’re out of credit or quota, or the model doesn’t exist, the chat now says so at once instead of retrying for up to 10 minutes. Errors quote the provider and say the one thing to do. Ollama and LM Studio say when a model isn’t downloaded, can’t use tools, or needs more memory than your PC has. Groq’s free plan can’t take a HyperDevs request, and the chat now says so.
  • • A lighter start. HyperDevs loads much less when it starts: the code for its first screen is 60% smaller, and dialogs and the generation tools load the first time you open them. The Viewer loads its 3D engine only when you look at a 3D model.
  • • A much smaller Studio plugin, installed first. The HyperDevs plugin for Studio is now 460 KB instead of 5.5 MB. HyperDevs installs or updates it as soon as it starts, so Open a game right after launch starts Studio with the plugin in place, and getting your project ready no longer waits for the AI engine. Restart Roblox Studio once after updating to load the new plugin.
  • • Failed generations explained. When a generation fails, the message now gives the reason in plain words and says your credits were refunded, instead of showing raw error text.
  • • When another program has HyperDevs’ port. If another program is using the port HyperDevs needs for the AI’s tools, a notice now says so, and the tools work again by themselves once the port is free.
  • • Help while Hyper 30 works. The chat now says how long a Hyper 30 build takes, and HyperDevs flashes its taskbar button and shows a notification when a long reply ends while you’re away. Closing HyperDevs during a reply asks first, and a very long message gets a tip to paste only what matters. The tour and the empty chat suggest asking for one thing at a time.

Fixed

  • • Optimize with every 3D model. Optimize now opens every model the 3D tab makes, Trellis 2’s included, and reaches the triangle count you pick on models it used to stop short on, such as Trellis 2 and Rodin v2. Optimizing again starts from the original model, not the last result, and a model already under the target is kept as it is, with a note. When something does go wrong, the message says what.
  • • Send to Game’s note in the chat. The note the Game chat gets after Send to Game showed up as a message from you. It’s now marked as a HyperDevs note, and the AI no longer reads it as something you asked.
  • • Models in Studio as they look in HyperDevs. Hunyuan3D models no longer go into Studio on their side, Rodin v2’s “Shaded” texture and models’ own vertex colours are kept instead of coming in grey, and every texture is fitted to what Roblox accepts. If a texture still won’t upload, the model goes in without it and the app says so.
  • • The model menu. It no longer gets stuck on “Loading models…”, and the chat always runs the model the menu shows. Ollama and LM Studio start on a model your server actually has, and the menu lists only models that can chat, without embedding, speech or image models.
  • • xAI. Chats on xAI’s Grok models failed every time. They now work.
  • • Long builds on your own key. A reply that reached its step limit on Claude, Gemini or Mistral ended in an error. It now ends with the usual note that you can send “continue”.
  • • Studio messages that said the wrong thing. Test Connection no longer says “Connected” with Studio closed. The Game tab no longer tells you to capture a window that comes in by itself, the Hierarchy no longer says to click a Refresh button that’s turned off, Open a game no longer asks you to close a place that isn’t in the Game tab, and the Studio plugin no longer tells you to close node.exe.
  • • Recovering from a window crash. In 0.1.30, a crash of HyperDevs’ window could close the whole app instead of reloading the window. It now reloads. If it crashes more than three times in a minute, HyperDevs asks whether to restart or quit.
  • • Applying a script change. While the AI’s script change is being written, Escape or a click outside the review no longer tells the AI you rejected it, and Ctrl+S no longer writes it twice.
  • • Stopping a generation. Stop didn’t cancel the job at fal.ai, so its credits were spent anyway. Stop now cancels it there. If fal.ai hadn’t started the job yet, your credits come back; once it has started, fal.ai bills it, so they stay spent. A message after Stop says which.
  • • Rigged characters. Asking the AI to rig a character failed every time, and your credits came back. Rigging now works, and a rigged character keeps the walk and run animations that come with it. Every clip of a character with several animations plays in the Viewer and the Animation library, not just the first, and the AI’s rigged characters preview with their animations. Saved animations keep their clips when you move your project folder, and deleting one deletes its clip files.
  • • The Restore button. The Maximize button now turns into Restore while the window is maximized.
  • • Two taskbar buttons. HyperDevs started from a pin no longer shows up as a second taskbar button. If a pin you made with 0.1.30 still does, unpin it and pin it again.
  • • Unfinished backups. Quitting while HyperDevs was backing up your project could leave a half-made copy that counted as one of your three backups. A copy now counts only once it’s whole, and an unfinished one is cleared at the next start.
  • • Smaller fixes. The Memory tab’s Patterns list can no longer squeeze its chat to nothing on a small window. The 3D tab’s detail setting says how many triangles the chosen model is asked for, and Optimize and Send to Game no longer tell you to turn on Studio betas that Studio no longer has. The MCP Tools list no longer names old image and 3D models, and z.ai now uses its standard API address unless you saved another.

Changed

  • • Micro-profiler capture. The AI can no longer capture Studio’s micro-profiler. The library it needed made up nine tenths of the Studio plugin.
  • • The Roblox Open Cloud key. It has moved from Settings → AI Providers to Settings → Integration, under Rigged character uploads, and you need it only to upload rigged characters from the Animation tab. There’s no Roblox User ID to type any more: HyperDevs asks Roblox which account the key belongs to and shows it.
  • • Hunyuan3D v3.1 Pro costs 90 credits. It was 50, less than each of its models costs to make. A 3D model now costs 10–90 credits, depending on the model, and the credit screens say so.
  • • Multi-view on Tripo H3.1. Multi-view now makes its model with Tripo H3.1 Multiview, at 40 credits, from the front, left, back and right views. Meshy v5, which it used before, is no longer available. Hyper3D Rodin is still there.

Security and privacy

  • • Further security hardening of accounts. We don’t publish the details.
  • • Usage data. The Privacy Policy now also lists how each Send to Game goes, when Studio connects, and how the Game tab shows Studio, and you’ll be asked to accept it once. Settings → “Share usage data” still turns it all off.

v0.1.30

October 16, 2026

Added

  • • Studio connects by itself. HyperDevs connects to Roblox Studio as soon as it starts, and keeps trying while Studio isn’t open, so there’s nothing to press. When a place opens, Studio’s window comes into the Game tab on its own, waiting until you bring HyperDevs back if it’s minimized; Settings can turn that off.
  • • Open a game. From the Game tab, open one of your recent Studio games, a fresh baseplate, or a place file, and Studio starts on it.
  • • Just the game view. The Game tab now shows only Studio’s 3D view, with its row of tabs, instead of the whole Studio window, since HyperDevs has its own Hierarchy, Properties and toolbar. It works whether the view is docked in Studio or floats in a window of its own. A switch in the Game tab’s header shows all of Studio whenever you need its menus or panels.
  • • Studio opens inside the Game tab. When you open a game from HyperDevs, Studio’s window comes into the Game tab right away and loads there, instead of covering your screen. Once the place is ready, the Game tab switches to the game view by itself.
  • • A new tour. Five short steps, offered once after you first sign in: connect Studio, build a floating island with one click, look at what the AI made, then a quick look at everything else HyperDevs does. Replay it any time with the new Get started button next to Settings.
  • • Rate a reply. A thumbs up or down under each finished reply, with an optional reason, tells us what works. It’s hidden when you’re signed out or have turned usage data off.
  • • Connect from the tab bar. Until Studio answers, the Studio light is a Connect button.
  • • Example prompts you can click. In an empty chat, click an example to put it in the message box.
  • • Credits about to expire. When some of your credits expire within 14 days, the balance in the title bar turns amber with an hourglass, and pointing at it says how many expire and when. Settings → Account now lists every expiry date, soonest first.

Improved

  • • Scrolling the chat. A chat you open now starts at the newest message, a queued message going out no longer pulls you away from what you’re reading, and when you’ve scrolled up, a Latest button takes you back down.
  • • Faster Hyper 30 replies. Every Hyper 30 request is about 2,700 tokens smaller, and replies use the model’s own recommended settings.
  • • Long Hyper 30 chats. A follow-up in a long chat can start answering sooner: when the chat allows it, Hyper 30 reads only what’s new instead of the whole conversation.
  • • No Studio setting needed. Studio’s “Allow HTTP Requests” doesn’t need to be on for HyperDevs, and new baseplates leave it off.
  • • Long Hyper 30 replies. A reply that’s still thinking is no longer cut off at 25 minutes. It keeps going for as long as the model makes progress, up to an hour. Memory’s scans and reviews now wait for a long think too.
  • • See what Hyper 30 is doing. While the model reads your request and thinks it through, the chat says so: “Reading your request…”, then “Thinking · 3 min”.
  • • When someone else is using Hyper 30. Your message now waits its turn and sends itself instead of failing, and a long reply keeps Hyper 30 until it’s finished.
  • • The Game tree catches up. After a reply that may have changed your game, the Game tab’s tree reloads by itself.
  • • The window reopens where you left it. HyperDevs now reopens maximized if you closed it maximized, and on the monitor you closed it on. If that monitor has been unplugged, it opens on your main screen, sized to fit. On a smaller laptop screen the window always fits, so the window buttons and the message box are never off the edge.
  • • Studio connection light. The light now shows whether Roblox Studio itself is connected: green while the HyperDevs plugin is talking to the app, yellow while the app is waiting for Studio. It no longer shows green with Studio closed, or yellow on a connection that is working. When Studio connects, the Game tab loads its hierarchy on its own.
  • • Model menu. The model menu now scrolls, and for providers with long lists — OpenAI, OpenRouter, Groq, Mistral — it has a search box and opens at the model you’re using. If your saved model has been retired, the app switches to the provider’s recommended model instead of the first name in the alphabet.
  • • Properties panel. A property changes only when you actually edit it. Clicking in and out of a field no longer rewrites it, an emptied box goes back to its value instead of setting it to 0, Escape cancels an edit, and text with line breaks gets a box that keeps them. A GUI object’s Size and Position get a UDim2 editor instead of a broken Vector3 one, and a property that used to get the wrong kind of editor, such as a ParticleEmitter’s Color, is now read-only.
  • • Panels remember how you left them. On every tab with a collapsible chat, a chat you collapse stays collapsed after a tab switch or a restart, a width you drag is kept, and the chat can no longer be squeezed into a thin strip. The Viewer keeps its layout too, and its preview panel collapses and reopens as it should. The Game tab’s split between Hierarchy and Properties and the Animation tab’s split between its view and controls are kept as well, and so is a divider you move with the keyboard.
  • • The Game tab’s Hierarchy. Folders you open stay open after a tab switch or a refresh, a click selects a row without closing it, the arrow keys start from the row you clicked, and a rename leaves the rest of the tree as it was. The toolbar’s Insert, Delete and Duplicate update the tree straight away and select the new object. With Studio closed, the Hierarchy says so and retries on its own.
  • • Chat history is saved as you go. A reply that finishes while you’re on another tab is saved at once, and closing the app saves anything still waiting.
  • • When the model server is busy. If a request is rate-limited, the chat now says so and counts down to the retry, instead of showing an empty reply. A provider that keeps retrying is no longer called unresponsive after two minutes; after ten minutes of retries, the reply ends with the provider’s own reason. If Hyper 30’s daily limit runs out, the chat says so and when it resets, instead of waiting and then saying nothing came back.
  • • Animation previews. With Loop off, an animation now holds its last pose and plays again when you press Play, and a model with several animations plays only the one you pick. The Viewer now plays animations too, and animated models there have Send to Animation.
  • • Notifications and Report a bug. Notifications now appear at the top right, clear of Roblox Studio and above open dialogs. “Report a bug” has moved to the title bar, and a report you were writing is kept if the dialog closes by accident.
  • • Settings shows what’s saved. In AI Providers, a provider reads “Configured” only once its key is saved, and an edit you haven’t saved says “Not saved”.
  • • Recovering from a crash. If HyperDevs’ window crashes, it reloads by itself, and the crash screen has window buttons. A tab’s Reload Tab brings back only that tab and keeps unsaved Coder edits.
  • • The Coder’s script list. Scripts inside other scripts are listed, each row shows the script’s parent, and with Trust agent on, the AI’s change to the open script shows in the editor right away.
  • • Hyper 30 on the generation tabs. Image Gen, 3D, Animation and Memory now send Hyper 30 a much shorter set of instructions, so it gets to your request sooner. On the Game tab, the AI playtests only when it changed a script or you asked it to.
  • • Send to Game from the 3D tab and the Viewer. It now opens the model’s folder straight away, with the file selected and ready to drag into Studio, and says so, instead of first waiting on a Studio import that couldn’t work.
  • • A lighter app. About 23,000 lines of code that nothing used are gone, including a background timer that ran every 30 seconds.
  • • Typing in the Coder. Typing in a script no longer redraws the script list and the chat with every key, so it stays quick in long scripts and long chats, even while a reply streams on another tab.
  • • Starting after an update. The first start after an update no longer waits while HyperDevs backs up your project. The backup now runs in the background once the app is up, so Studio and the AI connect sooner.
  • • The Memory map. Selecting something on the map no longer redraws every card, so it stays quick on big games.
  • • When another app has Studio. If another app already holds the connection to Studio, such as a second copy of HyperDevs or another Roblox Studio MCP server, HyperDevs now tells you which one, and connects by itself once that app closes.
  • • Long Hyper 30 prompts. When Hyper 30 takes a while to read a long chat, the request is no longer cut off and sent a second time, so it reads it once.
  • • Screenshots the AI can see. When you chat with a model that can see images, a screenshot of your game now reaches it. The toolbar’s Screenshot button copies the picture to your clipboard.

Fixed

  • • Ollama, LM Studio, OpenAI and z.ai models. Every chat on LM Studio, and most on Ollama, failed with advice to check an API key they don’t need, and some OpenAI and z.ai models you picked quietly ran the default model instead. The model you pick is now the one that answers, including one you add to Ollama or LM Studio while HyperDevs is open.
  • • Memory losing what it knew. A scan with Studio closed, or one that lost Studio part-way, no longer wipes what Memory already knew. Quick Scans, including the ones Auto-watch runs, no longer blank every script’s summary, and an AI Review no longer erases your patterns, confirmed systems and answered questions. An edit you make during a scan is kept. When the provider isn’t set up, or Hyper 30 is in use by someone else, a scan stops at once and says so, instead of saving placeholder summaries.
  • • Work lost on a tab switch. Switching tabs threw away the message you were typing, Image Gen’s prompt and image, the 3D tab’s source image and multi-view images, and the Viewer’s filter and search. All of it now stays, and anything that finishes while you’re away, such as an image the chat made, is there when you come back. The Memory tab’s Ask AI opens a collapsed chat and adds its question after what you’ve typed, instead of replacing it.
  • • Image Gen and the Viewer. Image Gen’s aspect-ratio buttons wrap instead of being cut off, a Viewer card’s buttons no longer also open its preview, and deleting an asset in the Viewer no longer leaves the 3D tab pointing at the deleted file.
  • • Studio covering dialogs. With Studio in the Game tab, moving or resizing HyperDevs while a dialog was open, or closing one of two open dialogs, could bring Studio up over the dialog. Studio now stays hidden while any dialog is open, and comes back in place when you restore HyperDevs after minimizing it.
  • • Links and long text in replies. Links in replies looked like plain text and did nothing when clicked; they now stand out and open in your browser. Long words and paths wrap instead of pushing the chat sideways, and code blocks always show as blocks you can scroll sideways.
  • • AI steps reported as failed. When a generation the AI started took over a minute, or a script change waited over five minutes for your review, the AI was told the step had failed while it was still going. It now gets the real result.
  • • Uploads to Roblox. With a Roblox Open Cloud key and User ID in Settings, the Animation tab’s Send to Game failed to upload a rigged model, and the Optimize tab’s Upload to Roblox failed to upload a texture. Both now upload, and when Roblox refuses one, you see its reason.
  • • 3D previews. The 3D, Optimize, Animation and Viewer tabs now frame every model the same way, so a large one no longer shows tiny, and Reset and the view buttons undo a pan. A model that fails to load in the 3D tab says so and can be tried again, a working preview no longer says it ran out of graphics memory, and a model with several materials frees its graphics memory when you move on.
  • • Keyboard shortcuts. The Game tab’s F5, Ctrl+Z and Ctrl+Y no longer act behind an open dialog, the Animation tab’s keys no longer scrub the animation while you’re in one of its dropdowns, and an Escape meant for a field or dialog no longer moves the Memory map up a level. Ctrl+1 to Ctrl+5 now switch tabs on AZERTY and similar keyboards, and Ctrl+F now searches a script you’re reading in the Coder, not only one you’re editing.
  • • The tab bar on smaller windows. Tab names could wrap onto two lines, and Settings and the status lights could be pushed off the edge. The bar now fits at every size the window allows: labels give way as it narrows, and at the narrowest the tabs show as icons that name themselves when you point at them.
  • • Updates that never finished. From this version on, an update download that stalls or is cut off ends with an error and Try again instead of hanging, and it keeps going if you leave Settings. Installing an update closes HyperDevs only once the installer has started, and the update screen’s download link opens the download page.
  • • Chats failing on retired models. Devstral, which was Mistral’s default, and DeepSeek V4 Flash have been retired, so chats on them failed with advice to check your API key. Mistral’s default is now Mistral Medium 3.5, and DeepSeek offers V4.1 Flash. Groq’s default is now GPT OSS 120B in place of the retired Llama 3.3 70B, and Gemini offers 3.6 Flash in place of 3.5 Flash.
  • • Studio actions that reported the wrong result. The Animation tab’s Send to Game no longer says “Failed to import model” after the model went in, and goes on to prepare the animation. Undo after applying the AI’s change to a script no longer says “Failed to undo” when it worked. The Game tab’s Screenshot says when Studio sent no image, and a failed Studio action now gives Studio’s reason.
  • • Chats and settings lost while saving. Rarely, a save could lose the chat, or your settings and API keys, if another program, such as a virus scanner or a backup tool, was holding the file at that moment. The saved copy now stays until the new one is in place, and a refused save is tried again. HyperDevs also waits for those saves when you quit.
  • • Tool changes that waited for a restart. Changes in the MCP Tools menu, including the ones a change of model makes, only reached the AI after a restart. They now apply at once, or when the reply that’s running ends, and a tool you turn off during a reply is off straight away.
  • • The AI missing the newest errors. When a playtest log was very long, the AI could miss the newest errors. It now gets the newest lines, and when any result from Studio is too long, the AI is told how to ask for less.
  • • Studio’s title bar after closing HyperDevs. Studio gets its title bar back when HyperDevs closes, even after a crash.
  • • A slow window list. The list of windows to bring into the Game tab now opens at once.
  • • “Check your API key” on Hyper 30. An empty Hyper 30 reply no longer tells you to check a key you don’t need; it says what happened.
  • • Saving Settings stopped a reply. Saving a provider in Settings no longer ends a reply in progress.
  • • Unsaved script edits lost. Going to another tab and back could replace your unsaved changes in the Coder tab with the version in Studio, and so could renaming the open script. Picking another script in the list discarded them without asking. Your edits now stay until you save or discard them, and switching scripts asks first.
  • • A keystroke in Properties flattening a script. Typing in a script’s Source row in the Game tab’s Properties could put the whole script on one line. Source is now shown there read-only; edit scripts in the Coder tab.
  • • Messages sent twice, or not at all. Pressing Enter twice quickly on Hyper 30 could send the same message twice, using two of your daily messages. A follow-up queued behind your last Hyper 30 message of the day was thrown away; it now stays in the queue. Messages queued during Hyper 30 maintenance now go out as soon as it ends.
  • • Optimize tab. After you loaded a second model, Save and Send to Game could export the first model’s optimized mesh under the second one’s name. Loading a model now starts the steps over, and a failed optimization no longer leaves the tab stuck on “Optimizing model…”. Send to Game no longer reports a preview for a build that never ran in Studio; it falls back to the clipboard. Apply Texture says which part got the texture, or that none was found. Send to Game no longer fails on models over 20,000 triangles: a big model goes in as several parts, and each mesh now sits inside its part.
  • • Deleted animations coming back. Deleting an animation from the Animation tab’s library now deletes its files, so it no longer reappears after a restart.
  • • Keys that closed or reloaded the app. Ctrl+W closed HyperDevs, Ctrl+R reloaded it — losing unsaved edits and stranding a reply in progress — and Ctrl+minus zoomed the page and put the Studio view in the wrong place. These browser shortcuts are gone, and a zoom left over from them is reset.
  • • A window that wouldn’t close. With unsaved edits in the Coder tab, the window could stop responding to Close after a reload, or on the update and crash screens. You’re now always asked, and can always close.
  • • The Studio permission prompt answered by accident. Keys you were typing in the chat — a letter, Enter or Escape — could answer the prompt that asks before the AI changes your game, and one accidental Deny blocked that action for the rest of the session. The prompt now takes only a deliberate click: Escape or the close button refuses just that request, Deny applies to that one action unless you tick Remember, and when the AI makes several changes at once, one Allow covers them all.
  • • Wrong prices on the credit screens. Buy Credits, Settings and Account said an image costs about 1 credit and a 3D model about 8. They now show the real range: an image costs 2–12 credits and a 3D model 10–60, depending on the model.
  • • Menus that lost focus. Typing in the provider search no longer loses focus after one letter, and an open menu no longer takes focus back while a reply is streaming.
  • • Changes landing on the wrong object. With two objects of the same name, the Game tab’s Properties, edits, Delete, Rename, Duplicate and drag acted on the first one. Each now acts on the object you picked. Names with spaces, hyphens or dots now work too, in the Game tab and the Coder.
  • • Studio selections cut down to one. A selection of several objects made in Studio is no longer reduced to one while the Game tab is open.
  • • A first message with no reply. The first message after installing or starting HyperDevs could end with “The model returned no response” before the model had even been asked. It now gets its answer, and sooner: HyperDevs gets the AI engine ready as it starts. If the engine still needs a moment, the chat says “Starting the AI engine…”.
  • • A reply that kept running. Reloading the window or clearing the chat now stops a reply in progress. Pressing Stop also closes a Studio permission prompt or a script change the reply had waiting for you, so neither can go ahead, and Studio actions or generations it had lined up don’t start.
  • • Replies cut off at the step limit now end with a note that you can send “continue”, instead of the engine’s own instructions.
  • • The AI reading positions, sizes and colours. When the AI read these from your game it got empty values; it now gets the real ones.
  • • “No MCP tools available” no longer shows in the chat header when HyperDevs starts before Studio.
  • • Buttons using the previous image or model. After a tab’s chat made an image or a 3D model, Make 3D, Download, Send to Game and the other buttons still used the one before.
  • • Models saved twice, and broken thumbnails. Leaving the 3D tab during a generation saved the model twice, and a model saved without a picture showed a broken thumbnail in the Viewer. The ✕ on an image you uploaded to the 3D tab now clears it.
  • • Memory tab. Editing a node’s summary could show, and save, another node’s text, and View Code could show another script’s code. On Windows, scripts under names with spaces, dots or hyphens had no saved code at all. Switching projects, or leaving the tab during a scan, left the old project’s details and no progress bar. A renamed project now keeps its Trust agent setting.
  • • Animation fields. Duration can now be typed below 1, a keyframe time left empty or out of range no longer breaks Send to Studio, and animation names with quotes now send correctly. Keyframe easing now uses Roblox’s pose easing styles; Sine, Quad and other styles a pose doesn’t have are eased as Cubic, and a notification says which.
  • • The tour offer is no longer declined by an Escape meant for something else, and no longer floats over other dialogs.
  • • Hyper 30 after you sign in on another computer. Signing in elsewhere replaces this computer’s Hyper 30 access, and your next message here failed with an engine error. HyperDevs now gets new access and sends the message again by itself.
  • • Two places open in Studio. With a second place open in Studio, every Studio feature and AI tool stopped working. They now go to the place in the Game tab, or else the one you opened last.
  • • Open a game after a Studio crash. A Studio that crashed no longer blocks Open a game for up to a minute; it frees up within about 20 seconds.
  • • Hierarchy search. Search found only objects in folders you had opened, and its count was wrong. It now searches your whole game, shows the first 50 matches without opening your folders, and counts the matches.
  • • Slow-loading playtests. A playtest that took a while to load was reported as a failed start, to you and to the AI, while Studio was already playing. It now says the game is still loading, for solo and multiplayer tests, and the toolbar’s Stop works.
  • • Studio switching places. When Studio moves to another place, a message waiting for Hyper 30 or queued for the old place is cancelled, with a note naming the new place, instead of going on in the new one.
  • • Two script changes at once. When the AI changed two scripts in one step, the second change’s review replaced the first, and the first failed five minutes later. The changes now come to you one at a time.
  • • 3D tab memory. The 3D tab now frees a model’s graphics memory once you move on from it.
  • • Space on the Animation tab’s buttons now presses the focused button, as Enter does, instead of playing or pausing.
  • • Optimize files. Picking the same file again now loads it, so a fresh export from Blender comes in. A “.gltf” save now writes real glTF that other tools open, and a “.gltf” that needs files beside it says which ones. When Send to Game publishes a mesh but can’t build its part, the message names the published asset.
  • • The Coder’s Edit button now puts the cursor in the script, so you can start typing.
  • • Script changes reported as failed. Applying the AI’s change could say it failed, and tell the AI so, when Studio had it and only reloading the editor went wrong. When the AI tries to change a script that doesn’t exist, it’s now told to create it first, instead of showing you a change that can’t be applied.
  • • Deleted assets left on disk. Deleting an asset removed it from the library but left its files on your disk. The files are now deleted too.
  • • Replies waiting on a hidden question. The AI could ask a question HyperDevs never showed you, and the reply then waited until it timed out. It can no longer ask one.
  • • MCP Setup Guide. When Windows can’t open the guide, Settings now says why and where the guide is, instead of doing nothing.
  • • Saved animations. An animation saved to your library now keeps its animation files with it, so it opens with its animation in the Viewer and the Animation library.
  • • Smaller fixes. The message box shrinks back after you send a long message, its placeholder fits on one line, and your messages show one avatar, not two. Old test results and errors no longer linger in Settings, the window picker and the Buy Credits dialog. The tour and Settings point at the right places, the tour’s arrows stay on their targets when you switch tabs or drag a divider, and the tabs’ tooltips no longer name old models. Right-click menus close when you scroll, the Animation tab’s Export section folds, and Open Settings on the chat’s “API Key Required” notice goes straight to AI Providers.

Changed

  • • Memory scans on Hyper 30. Scans no longer run while Hyper 30 is under maintenance, and they have a daily allowance of their own, so a big scan can’t shut you out of Hyper 30 chat for the rest of the day. A scan that stops keeps its work and says why, and when a limit resets. Scans on another provider aren’t limited this way.
  • • Duplicate in the Animation library. The Duplicate button has been removed. It only ever copied an animation until the app closed.

Security and privacy

  • • Further hardening of the AI’s Studio permissions. We don’t publish the details.
  • • Further security hardening of accounts and of the app. We don’t publish the details.
  • • Account suspensions are enforced more reliably.
  • • Where Hyper 30 runs. Settings → AI Providers now lists Hyper 30 in its own group, “From HyperDevs”, instead of under “On your machine”: Hyper 30 messages are processed on HyperDevs’ servers, while Ollama and LM Studio keep everything on your computer.
  • • Usage data and reply ratings. The Privacy Policy now lists everything the app records, ratings included, and you’ll be asked to accept the updated terms once. Settings → “Share usage data” turns it off.
  • • Chats stay with their account. Each account’s chats are now kept apart, so on a shared computer you see only your own. If you’re signed out without choosing to be, the app says so, and your chats come back when you sign in again.
  • • The AI engine stays quiet. It no longer downloads anything on its own; everything it needs now comes with HyperDevs.
  • • Further privacy hardening of the AI engine and of the Studio plugin. We don’t publish the details.
  • • We cleaned older records of data we no longer need.

v0.1.24

September 2026

Added

  • • See what the AI is writing. While the AI writes a step — often a whole script — the activity line now names it, for example “Writing set_script_source…” and then “Running…”, instead of staying on “Thinking…” for minutes.
  • • Ollama and LM Studio check. If Ollama or LM Studio is selected but isn’t running, or has no model loaded, the chat says so as soon as you pick it and tells you what to start, instead of sending a message that can only fail. If the server is running but lacks the selected model, the app switches to one it has.

Improved

  • • Faster Hyper 30 replies. Each message now makes one request to the model instead of two, and every request carries a shorter description of the Studio tools, leaving more room for your project.
  • • Precise script edits. When the AI reads a script it now sees line numbers, so it can change the lines it means to instead of rewriting the whole script.
  • • Opens where you left off. HyperDevs now opens straight to the tab you last used, instead of loading the Game tab first.
  • • Accurate status lights. The Studio connection light stays accurate on every tab, not only while the Game tab is open, and the LLM light now shows whether the AI engine is running. The “LLM provider ready” pop-up no longer appears on every launch.
  • • Resizable Patterns list. In the Memory tab, drag the bottom edge of the Patterns list to make it taller or shorter. The height is remembered.
  • • Optimize. The two most common optimization errors now suggest the right fix, and the chat panel keeps the size you give it.

Fixed

  • • “Sent nothing back for 6 minutes” during long thinking. Hyper 30 could show this error while it was still thinking through a big request. The check now counts thinking and tool writing as progress, and a long response is no longer cut off at 10 minutes.
  • • Hyper 30 slowing down after a failed message. A message the app had given up on could keep running on the Hyper 30 server and hold up the messages sent after it. It is now stopped properly.
  • • Failed messages using up the daily allowance. A Hyper 30 message now counts toward your 15 per day only once the AI starts answering, so a message that fails while thinking doesn’t use one up.
  • • Hyper 30 access being renewed over and over. The app could set up Hyper 30 access several times in a row, restarting the AI engine each time and interrupting a response in progress.
  • • Settings that could be lost. Rarely, two settings saved at the same moment could discard your other settings, including API keys. Settings are now saved one change at a time.
  • • Send to Optimize and Send to Animation. Sending a model from the Viewer or the 3D tab could open the target tab empty. The model now arrives loaded.
  • • Memory tab. Re-scanning a single script can no longer clash with a scan already running, and filtering the map keeps every folder above a match.
  • • Animation and Viewer. Playback no longer jumps forward after a pause, Shift+L no longer also toggles looping, the Keyboard Shortcuts section opens, and keyframe edits always register.
  • • Smaller fixes. Crash screens for the Animation and Memory tabs name the tab instead of saying “Unknown”, the tour offer no longer hides the Studio view, and a model that fails to save no longer leaves its files behind.

Security and privacy

  • • Further hardening of the AI’s Studio tools. We don’t publish the details.
  • • Usage data. The app now notes when a reply finishes and how long it took, so slow or failed replies can be found and fixed — never what the reply said.

v0.1.23

September 2026

Added

  • • Live thinking. When a model reasons before it answers, that reasoning now appears as it is written instead of arriving in one block once the model has finished. On a slow first response you can see the work happening rather than watching an empty spinner.
  • • Close confirmation. Closing the window with unsaved work now asks first — from the title bar, the taskbar or Alt+F4 — and choosing “Keep editing” genuinely keeps the app open.

Improved

  • • A much smaller download. The installer is about 149 MB, down from about 509 MB, and a full installation takes roughly 562 MB instead of 1.1 GB. Nothing was removed to get there — the build was shipping files it never used.
  • • Smoother chat. A long conversation no longer redraws every message on each word the AI streams, so replies stay responsive as the conversation grows.
  • • 3D model picker. The 3D tab lists the models actually on offer, with prices in credits, and defaults to Trellis 2. Multi-view prices come from the same table, so the figure you are quoted is the figure you are charged.
  • • Clearer messages. Failures from the generation service are described in plain sentences, and a suspended account is told so instead of being shown a confusing daily-limit message.
  • • Accurate setup guidance. The tour, Settings and the in-app guides now describe the Studio plugin that ships with the app, rather than a manual installation that has not been needed for some time.

Fixed

  • • Sign-in appearing to hang. Entering the right password could leave the dialog spinning even though you were already signed in — closing and reopening the app showed you signed in all along. Signing in no longer waits on your profile loading.
  • • “Setting up Hyper 30 access” that never cleared. The message could stay on screen indefinitely, and signing out and back in — the very thing it suggested — could be what caused it. Switching accounts now works without leaving Hyper 30 stranded.
  • • Silently shortened scripts. A script or property longer than the internal limit was quietly cut off. Over-long values are now refused with an explanation, so nothing is truncated without you knowing.
  • • The AI is told when a Studio action fails. Failed operations could be reported back to the AI as successes, so it carried on as though the change had been made. Failures are now passed through as failures.
  • • Create New Script. Creating a script now returns its location, and the toolbar selects the script it just inserted.
  • • Buttons that did nothing. The Save to Library button in the Animation tab never worked and has been removed, as has a subscription plan that was never available to buy.
  • • Smaller interface fixes. Error indicators no longer disappear from failed messages, a project opened from the Studio plugin is remembered, and notifications are drawn once rather than twice.

Changed

  • • The Hyper 30 model. Hyper 30 now runs a dense Qwen3.8 27B. Replies may read differently from previous versions, and this is why.
  • • Windows on Arm. HyperDevs now ships a single 64-bit Intel/AMD build. The separate Arm build could not load part of the image pipeline and ran the Intel components regardless; Windows on Arm runs this build through its built-in emulation.

Security and privacy

  • • Further hardening across the app’s local connections and the services behind Hyper 30, generation and payments. We don’t publish the details.
  • • Less in your usage data. Use of the app while signed out is no longer recorded, and what is recorded is only ever uploaded for the account it belongs to.
  • • Less in your bug reports. Email addresses and Windows user-folder names are removed from the diagnostics attached to a report, and no text from the chat, the editor or form fields is included.
  • • Credits are returned when generation fails. A generation charged and then unable to finish is refunded, and refunded credits keep their original expiry date.

v0.1.22

September 2026

Added

  • • Email verification at sign-up. New accounts are confirmed with a 6-digit code sent by email. The sign-up form asks for your email address twice to catch typos, lets you request a new code or correct the address, and takes you to Sign In if the address already has an account.
  • • Password reset. Select “Forgot password?” on the sign-in screen to receive a 6-digit code by email, then choose a new password in the app. The reset email contains the code only — no links.
  • • Account deletion. Delete your account from Settings → Account. Deletion is permanent, requires your password, and removes your profile, credits, purchase history, generation history and usage data. Projects and files on your computer are not affected, and any bug reports you submitted are kept with your account details and diagnostics removed.
  • • Daily Hyper 30 allowance. The Free plan includes 15 Hyper 30 messages per day. Only the messages you send are counted — not the individual steps the AI takes to complete them — and a message counts only once Hyper 30 begins responding. Providers you use with your own API key are not limited.
  • • Usage data setting. A new “Share anonymous usage data” option in Settings → General. When it is on (the default for signed-in users), HyperDevs records which screens you open, your progress through the tour, and when requests fail — never your prompts, code, file paths or generated content. Turning it off also discards anything not yet sent.

Improved

  • • Automatic Hyper 30 setup. Hyper 30 access is provisioned when you sign in, with nothing to request or configure. If setup is still in progress, the app says so instead of failing silently.
  • • Elapsed-time indicator. The activity tray now shows how long the AI has been working (for example 12s, then 1:23), making it easier to tell a slow request from a stalled one.
  • • Clearer account messages. Sign-in, sign-up and verification errors are now described in plain language, including how long to wait before requesting another email.

Fixed

  • • Hyper 30 on new installations. New installations could not use Hyper 30: removing invite codes had also removed the only step that issued an access key, and the app showed no warning. Access is now provisioned automatically.
  • • Credit purchase confirmation. The app could report that credits had been added before the payment was processed. It now waits for your balance to update before confirming.
  • • Links from the browser. Payment and email-confirmation links now open the app to the right screen. Previously the app launched but ignored the hyperdevs:// link.
  • • Tool settings. A single malformed value could reset all tool settings and disable the input-simulation tools. Your settings are now preserved.
  • • Hyper 30 queue. The single-GPU queue now identifies you by account rather than by device, so signing in on another computer no longer affects your place.
  • • Hyper 30 while signed out. Hyper 30 remained available after signing out, outside the daily allowance. It now requires a signed-in account, and signing out removes Hyper 30 access on that device.
  • • Signed-out state after a connection drop. If HyperDevs could not reach its servers at launch, it showed you as signed out even after the connection recovered. It now restores your session automatically.
  • • AI engine restart at launch. Setting up Hyper 30 while the AI engine was still starting could stop the engine until the app was restarted (“The AI engine failed to restart”). The engine now finishes starting before the change is applied.

Security

  • • Hyper 30 queue management functions now require a signed-in session.
  • • Deleting an account requires re-entering its password, so an unattended, signed-in computer cannot be used to delete it.
  • • Signed-out clients no longer make failing authentication requests during background checks.

v0.1.21

September 2026

Added

  • • Open beta. Invite codes are no longer required — accept the Terms to get started.
  • • Code editing in the Coder tab. The script viewer is now a full editor with Luau syntax highlighting that saves directly to your live Studio session. You can also create, rename and delete scripts. Before saving, the editor re-reads the script and warns you if it changed in Studio, rather than overwriting it.
  • • Guided tour. A 14-step walkthrough on first launch. It can be skipped, and replayed at any time from Settings → General.
  • • Light and dark themes. Switch themes from the title bar. Dark remains the default.
  • • Relations view in Memory. See how your scripts connect — which scripts fire which events and require which modules — with each link marked as confirmed, runtime or dangling, and a filter to isolate broken links. Rebuilding the view takes seconds; previously it required a 20-minute deep scan.
  • • Patterns panel in Memory. Coding patterns identified by the AI review are now displayed in the app.

Improved

  • • Better project awareness. The project summary the AI receives before each task now names your scripts, grouped by system. On a 43-script test project, scripts named rose from 0 to 43, patterns supplied rose from 8 to 25, and useful context grew from 961 to 2,472 tokens.
  • • More coherent long conversations. Chat history is now budgeted, and the AI is told when older messages are left out, which stops it contradicting earlier parts of a long session.
  • • Faster chat loading. Long histories open faster: the most recent 60 messages are shown first, with earlier messages available on request.
  • • More accurate Memory search. Results are now ranked by relevance; previously every summarized script matched every query.
  • • Lower background CPU usage. Background checks pause while the window is hidden, including a Studio round-trip that previously ran about 50 times a minute.
  • • Refined visual design. A neutral background replaces the full-screen gradient, with one button style and consistent hover states throughout. The Store and Apps tabs now use the standard background.

Fixed

  • • Scripts containing backslashes. Escape sequences such as \n in strings, and Windows paths such as C:\temp, were altered when scripts were sent to Studio.
  • • Memory reference map accuracy. Script-relative references (script.Parent.Something) were not resolved, producing entries for instances that do not exist while coverage was reported as complete. On a test project, dangling links fell from 6 to 0, and references that cannot be resolved are now reported accurately.
  • • WaitForChild with a timeout. WaitForChild("Event", 5) now resolves correctly in the reference map.
  • • Rename instance. Renaming from the Game tab's context menu now works; it previously failed silently.
  • • Create New Script. Creating a new script now works as intended.
  • • Stopping jobs. Stopping 3D, animation and rigging jobs now takes effect within a second, down from up to five.
  • • Theme flash at launch. The app no longer flashes through several backgrounds while starting.
  • • Optimize vertex merging. A mesh-welding setting ignored since a library upgrade is applied again, restoring the expected vertex merging.
  • • Reading script source. The AI can read a script's source on the standard path again.
  • • Memory auto-scan. Scheduled scans every five minutes now run regardless of which tab is open.
  • • Unverified badge. Memory nodes can now be marked as verified, which clears the “Unverified” badge.
  • • Also fixed: Windows confirmation dialogs replaced with in-app dialogs, duplicate chat messages when several tool calls completed at once, and the local model starting a playtest unprompted.

Security & reliability

  • • General security and stability improvements.

v0.1.20

August 2, 2026

Added

  • • New deterministic 2-step Rig pipeline — reliably rig and animate 3D models (the old one-shot path sometimes returned a static mesh for non-humanoid subjects)
  • • 5 new image models (Flux Pro 1.1 + Ultra, Flux 2 Pro, Ideogram v3 / Turbo / Quality) and 4 new 3D models (Hunyuan3D v3.1 Pro with real textures, Trellis 2, Tripo H3.1, Hyper3D Rodin v2)
  • • Customizable Game-tab toolbar

Improved

  • • Noticeably lighter and faster — removed ~9,400 lines of dead code and swapped in a far smaller code highlighter
  • • Optimize no longer freezes the app — mesh decimation now runs on a background thread
  • • Smoother chat scrolling while streaming and lower CPU use while the Studio overlay is active

Fixed

  • • In-app updates now actually download — the updater previously stalled at "Downloading…" forever and could never complete, so updates had to be installed by hand
  • • 3D models now render correctly in the installed app — compressed meshes (Rodin, Meshy, Hunyuan3D) were silently failing to decode
  • • The local AI engine now starts reliably on a fresh install
  • • Credit billing hardened — closed races where a cancelled or racing job could hand over an asset and refund the credits
  • • Meshy v6 animation no longer charges and reports success when it can't rig a subject — you now get a clear message and a pointer to the Rig pipeline
  • • Plus: animation source-image quality, false "Purchase Successful" toasts, sign-out state cleanup, fetch hangs, and Luau injection guards

Security & reliability

  • • Hardened authentication on the local Roblox Studio connection
  • • Stop now reliably cancels in-flight generations server-side and refunds credits
  • • API keys and secrets remain encrypted at rest
  • • Updates are verified end-to-end — installer files are checked against a published checksum before they run, and anything that doesn't match is refused
  • • Strengthened authentication for generation callbacks
  • • Content-safety screening now covers every prompt field, with provider safety filters enforced server-side
  • • Out-of-date installs now show a clear "update required" prompt instead of failing with confusing errors
  • • Updated image-processing and packaging dependencies to pick up upstream security fixes

v0.1.15

July 1, 2026

Added

  • • 3D and Animation generations now keep running when you switch tabs and finish when you return — previously leaving a tab cancelled an in-progress generation
  • • "Clear" button on the 3D Models action bar to discard the current model and start fresh
  • • "Show in Folder" and "Copy Path" on asset cards and the Viewer for easy manual import into Studio

Improved

  • • Sharper Flux Dev images (PNG output) and correct multi-view source-image quality
  • • Per-model parameter handling across the image and 3D generation pipeline for more reliable results

Fixed

  • • 3D models from Rodin, Meshy, and Hunyuan3D (compressed meshes) no longer crash the viewer — the content security policy now permits the WebAssembly mesh decoders they require
  • • The cascading "Session timed out" error that could appear after a 3D generation (this was downstream of the viewer crash above)

Security & reliability

  • • The local Studio bridge now requires authentication, closing a path that could have allowed remote code execution on your machine
  • • Hardened the HTTP layer against request-smuggling and server-side request forgery (SSRF) attacks
  • • Fixed path-traversal vectors that could have let file access escape its intended directories
  • • The Stop button now reliably cancels in-flight generations and cleans up orphaned background processes
  • • More reliable Memory scanning with fewer missed files and better error recovery

v0.1.14

June 18, 2026

Added

  • • Roblox Studio MCP now runs in-process and the Studio plugin auto-installs — no Node/npx and no manual plugin setup
  • • "Hyper 30" built-in cloud model — generate assets and chat with no API keys of your own (credit-billed, with a free tier)
  • • Per-model tool presets: enable or disable MCP tools per model from any chat panel
  • • Rebuilt Memory system — graph, blob, and AST indexing of your project for richer AI context, with Tree, Graph, and Context views
  • • Expanded generation suite: 6 image models (Flux Schnell, Flux Dev, Flux Pro, Flux 2 Pro, SDXL, Ideogram) and 5 image-to-3D models (TripoSR, Tripo3D v2.5, Trellis, Hyper3D Rodin, Hunyuan3D)
  • • OpenCode powers the local LLM harness behind the chat

Improved

  • • Redesigned generation pipeline with clearer progress and real error surfacing
  • • Hardened credit and usage endpoints, and a more reliable Studio connection flow

Fixed

  • • Numerous stability fixes from the 0.1.7–0.1.13 testing rounds (new-user install path, coder/overlay/LLM issues, AI tool guidance)

v0.1.7 - v0.1.13

June 2026

Improved

  • • Iterative stability and performance work across the closed-beta testing rounds — incremental refinements to startup, panel layout, the Coder/overlay/LLM stack, and asset-generation reliability as feedback came in

Fixed

  • • Numerous bug fixes and regressions addressed between v0.1.6 and v0.1.14 — including the new-user install path, overlay and LLM issues, and AI tool-guidance refinements. These fixes were rolled up and shipped in v0.1.14.

v0.1.6

June 2, 2026

Added

  • • HyperDevs Cloud (beta): authenticated cloud generation with a free daily credit tier — get started without your own API keys
  • • llama.cpp support as a local LLM provider, alongside Ollama
  • • Redesigned Settings with sidebar navigation and HMAC key rotation
  • • Animation timeline scrubbing to preview generated animations

Improved

  • • Major performance pass — removed background polling, optimized React rendering, and lazy-loaded heavy libraries for faster startup and lower memory use
  • • Accessibility — screen-reader (ARIA) labels added to loading states across every tab
  • • More robust asset generation pipeline with clearer progress and error handling

Fixed

  • • Animation tab export-panel and multi-view progress bugs
  • • Memory leaks in the tool logger and 3D model viewer
  • • Race conditions in panel-layout persistence and window capture
  • • Crash-safe (atomic) writes for stored keys to prevent corruption

Security

  • • Hardened credit/usage endpoints and rate-limit checks
  • • Removed API-key fragments from logs
  • • Electron security updates for known CVEs

v0.1.0

Initial Release May 5, 2026

Added

  • • Initial release of HyperDevs
  • • 8 AI-powered tabs: Game, Image Gen, Coder, 3D Models, Animation, Optimize, Memory, Viewer
  • • Memory tab with game knowledge graph
  • • Viewer tab for asset preview
  • • MCP integration with Roblox Studio
  • • Support for 5 AI providers: Claude, Gemini, OpenAI, z.ai, Ollama
  • • Text-to-image generation with Flux Schnell
  • • Image-to-3D with TripoSR and Tripo3D
  • • AI animation generation with Meshy v6
  • • Mesh optimization pipeline for Roblox limits
  • • One-click import to Roblox Studio via EditableMesh
  • • Multi-view image generation
  • • Local LLM support via Ollama
  • • Encrypted local API key storage

v0.2.0

Coming Soon

More features and improvements on the way — get in touch to be notified when updates are released.

Stay Updated

Want to be notified about new releases and features? Get in touch.

Contact us →