Instructions to use deskmind/brain-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use deskmind/brain-4b with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir brain-4b deskmind/brain-4b
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
DeskMind Brain 4b · 得心
得心,应手。 Typed, calibrated step decisions for computer-use agents, running locally on Apple Silicon with MLX. This model is the strong tier (answers the steps the router escalates) of the DeskMind Brain router.
- Release:
g14-q8(checkpoint G14, 8-bit MLX, prompt format 3). Earlier releases are on branchesg13-q8andv7b-q8. - Code and usage: deskmind-ai/brain
- Prompt format:
deskmind.jsonrecords the format this model was trained with. Keep it next to the weights.
Use
uv run hf download deskmind/brain-0.8b --local-dir models/brain-0.8b
uv run hf download deskmind/brain-4b --local-dir models/brain-4b
uv run deskmind-brain-serve --predictor mlx:models/brain-0.8b --escalate-to mlx:models/brain-4b --two-stage
Results (this 8-bit release, served as shipped)
| DeskMind router (0.8B → 4B) | Jev (TypeSafe AI, cloud reference) | |
|---|---|---|
| Real macOS desktop, 13 sandbox tasks × 3 runs (bench v23) | 92% (35/38 scored), 0 false "done" | 87% (33/38 scored), 2 false "done" |
| Time per step (p50 / p95, M4 Pro) | 0.59 s / 4.6 s (local) | 0.36 s / 0.44 s (hosted API) |
Known limits:
- Slower than cloud models.
- It fails a web-extraction task, which the reference model also fails. When a request is ambiguous, it asks the user before acting.
- It was trained on macOS Finder and TextEdit sandbox tasks plus web and form decisions, and is untested elsewhere.
Training
LoRA distillation on Qwen/Qwen3.5-4B (KL to teacher distributions plus cross-entropy to labels), merged and quantized to 8 bits. The data is desktop DAgger in sandboxed tasks labelled by an oracle, plus synthetic and public web and form decision data. It contains no real user data. Details: docs/training.md.
中文: 得心(DeskMind)Brain 的强档(路由交上来的步骤由它回答)。给电脑操作 agent 的每一步做带类型、带把握程度的决策,用 MLX 在 Apple Silicon 本地运行。
- 真机成绩: 13 个沙箱任务各跑 3 轮,通过率 92%,同条件下 Jev 为 87%。
- 速度: 每步中位约 0.6 秒。
- 详见 deskmind-ai/brain。
License
Apache-2.0. Fine-tuned from Qwen/Qwen3.5-4B (Copyright Alibaba Cloud, Apache-2.0). The DeskMind name, 得心 and the logo are not covered by this licence.
- Downloads last month
- 2
8-bit