Cerebellum-2B
Open 2B non-autoregressive decision model on Qwen3.5-2B for agent tool routing and DOM actions, with BF16, FP8 and INT8 weights.
Overview
Cerebellum-2B pitches itself as the cerebellum to an LLM's cerebrum: a System 1 decision engine for high-frequency agent actions such as tool selection and DOM automation. Built on Qwen3.5-2B, it picks an action from a candidate set in one forward pass without KV cache growth; the README cites 25 ms inference and about 40 ms locally on a 16 GB M4 MacBook with the INT8 build.
Weights come in BF16 (3.76 GB), FP8 (2.39 GB, for vLLM servers) and INT8 (2.23 GB, for CPU and edge), with small reported drift from quantization. The README claims 94.92% accuracy on agent routing, ahead of Laya, Kev, Jev and GPT-4o structured output on its own evaluation of more than 95,000 samples across six scenarios; these are the author's figures. Docs are in Chinese and English with an architecture white paper.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which tool or DOM action should the agent take from these candidates? | candidate tools or actions |
How it works
- Agent presents state and candidate actions
- Model scores candidates in one forward pass
- Pick the top action
- LLM handles planning and reflection
Strengths
- Very low latency
- Quantized builds for edge devices
- Apache-2.0 weights
Caveats
- Benchmark claims are self-reported
- Marketing-heavy README
Getting started
Download weights from Hugging Face (mkzero/Cerebellum-2B-BF16, -FP8 or -INT8).At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- vLLM / FastAPI / CPU
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Non-Autoregressive AI Agent Decision Model. Open-source SOTA alternative to TypeSafe Jev. O(1) Tool Routing & DOM Automation on Qwen3.5-2B .
hellogumbo/awesome-jev
Frequently asked questions
What does it decide?
Which tool or DOM action an agent should take from a candidate set.
Are the SOTA claims verified?
They come from the author's own evaluation.
Can it run on a laptop?
The INT8 build is aimed at CPUs and a 16 GB M4 MacBook.
Related projects
Local browser agent that picks each step's operation and element with a Jev-shaped model such as a fine-tuned Laya, no cloud needed.
MCP serverbetaPythonComputer-use platform (Cua Spaces, Driver, Lume) with CUA-S1 small decision models and a preview jev-use example for bounded actions.
ProjectbetaPythonStdlib Python dispatcher turning typed function signatures into one batched Jev request, with a browser harness against ground truth.
ProjectprototypePythonOpen 2B browser-agent decision model fine-tuned from MiniCPM5 that serves the System One protocol for jev-ultrafast.
ProjectresearchPythonLangGraph dental-clinic agent where a local Jev-style router picks which tools to bind each turn, halving LLM token use.
LibraryprototypePythonOpen Apache-2.0 browser-agent decision model (Qwen3.5-35B-A3B fine-tune) that doubled Jev 1.13's task success in the same agent.
ServiceresearchPython