Advanced

Power-user techniques for getting the most out of HyperDevs.

Custom System Prompts

You can customize how AI responds by setting custom system prompts for each tab.

Use Cases

  • • Coding Style: Enforce specific patterns or frameworks
  • • Response Format: Request JSON, Markdown, or other formats
  • • Domain Knowledge: Add game-specific context or rules
  • • Constraints: Set limits on response length or complexity

Understanding EditableMesh Limits

EditableMesh is powerful but has limits. Understanding these helps you create better assets.

Key Limits

Triangles per part ~10K recommended
Texture size 1024x1024 or lower
Materials per mesh 1 (baked recommended)

Send to Game imports through EditableMesh. The 3D Models tab and the Viewer reduce the model to about 10,000 triangles first; from the Optimize tab, a model over 20,000 triangles goes in as several parts, each mesh inside its part.

Hyper 30 & Local Models

Three ways to run AI without a third-party API key: the built-in Hyper 30 model, or a fully local model via Ollama or LM Studio.

Hyper 30 (built-in, no setup)

The default model. It runs in the cloud, through HyperDevs' proxy, while Ollama and LM Studio keep everything on your computer. Just sign in and it works — free to use every day with an account, and you can keep going with credits after it.

  1. 1. Sign in with your HyperDevs account
  2. 2. Find it under Settings → AI Providers → From HyperDevs. There's nothing to enter: access comes with your account
  3. 3. Start chatting — no key needed. Image, 3D and animation generation use credits

What to Expect

  • • Free use: a meter under the message box shows how much of today's free use you've used. It resets at 00:00 UTC, and a message you start with free use left finishes free, within limits
  • • Long replies: a reply can keep thinking for up to an hour. The chat shows "Reading your request…", then "Thinking · 3 min"
  • • After free use: the chat says so and when it resets, and offers to keep going with credits. That's off until you turn it on, with a daily cap you set, and you can turn it off any time in Settings → Account
  • • Memory scans: scans on Hyper 30 use your free daily use and never credits. They stop when it's gone, and don't run during Hyper 30 maintenance

Ollama (free, fully offline)

Run models locally for complete privacy and offline use.

  1. 1. Download Ollama from ollama.com
  2. 2. Install and run ollama serve
  3. 3. Pull a model: ollama pull gemma4:26b, or ollama pull llama3.2 for a lighter one
  4. 4. In HyperDevs, select Ollama as your provider (default: http://localhost:11434)

Recommended Local Models

  • • gemma4:26b: App default — best balance of quality and speed
  • • llama3.2: Lightweight and fast

LM Studio (local, no key)

Another way to run models on your own computer. HyperDevs uses whatever model LM Studio has loaded.

  1. 1. Install LM Studio and load a model in it
  2. 2. Start LM Studio's server from its Developer tab
  3. 3. In HyperDevs, select LM Studio as your provider (default: http://localhost:1234)

If Ollama or LM Studio isn't running, or has no model loaded, the chat says so as soon as you pick it and tells you what to start.

Studio's Built-in MCP Server (Experimental)

Most people use the default connection: the HyperDevs Studio plugin, which HyperDevs installs for you and which connects by itself. As an experimental option, Settings → Integration can instead connect through the MCP server Roblox now builds into Studio. The option shows only with "Show experimental features" on, and needs Studio's own "Enable Studio as MCP server" switch, found in the Assistant panel’s … menu (Roblox’s docs call the page Manage MCP Servers; in some Studio versions it’s under Settings → MCP Servers).

Documentation for HyperDevs v0.1.34 Report an issue