Local AI RecipesLocal LLM / Agent

MiniCPM5-2B local Agent recipe

A compact Apache-2.0 local model for tool-use experiments — with a very important warning about giving any model control of a real desktop.

Commercial-friendly model · Safety-limited productReviewed 2026-09-26
Commercial statusMiniCPM5 models are released under Apache-2.0. Check the runtime and any extra tools you add separately.
HardwareDesigned for on-device and resource-constrained deployment. Start small; choose quantization and context size for your actual RAM/VRAM.
Runtimellama.cpp / GGUF or other compatible local runtimes; Ollama is a convenient DIY path when a compatible model package is available.

The interesting part is not installing a 2B model. Tools such as Ollama already make model installation easy. The interesting part is what happens after the model can call files, terminals, browsers and desktop controls.

AILocalabs therefore recommends MiniCPM5-2B as a local experimentation recipe, but does not currently ship a general-purpose desktop Agent for ordinary users. We are not satisfied with the host-computer safety boundary yet.

DIY recipe

Build it yourself

  1. Run the model through a local runtime such as llama.cpp/GGUF or a compatible Ollama package.
  2. Give the Agent structured tools first: file.read, file.write inside an allowed workspace, git.status, build/test and explicit browser actions.
  3. Keep unrestricted shell and whole-desktop control disabled by default. Add them only in an explicit developer mode.
  4. Treat web pages and retrieved documents as untrusted data. Prompt text must never grant new machine capabilities.
Risks

What can go wrong

  • A non-admin process can still delete personal files, upload data or destroy a project.
  • Prompt injection can turn browser content into unintended tool calls.
  • Desktop automation can bypass a shell policy by opening a terminal and typing commands.
  • A consumer product needs a capability broker, scoped folders and clear confirmation for side effects.
For humans and AI agents

Give the recipe to your AI and reproduce it locally

Target hardware, runtime, key parameters, validation and risks are published as Markdown and JSON. An AI agent should still inspect your machine first, explain changes and ask before risky operations.

AILocalabs decision

What we ship — and what we do not

Recipe only for now. Luna Unlimited remains a developer/internal capability; we will not sell a general desktop Agent until Safe Mode is useful enough on its own.

Upstream / license sourcesMiniCPM official repository
Recipe discussion

Tried it? Leave results, questions and improvements.

Discussion stays with this recipe so future builders can learn from it. Never post passwords, API keys, license keys or private files.

Loading discussion…

Join the discussion