DeskMind Brain 4B
The decision model of DeskMind, an open-source computer-use agent for macOS: Qwen3.5-4B with a merged LoRA, 8-bit MLX, answering choice, yes/no and score questions from answer-letter logits. Its server speaks /v1/systemone on Apple Silicon; a 0.8B sibling is the fast tier.
A personal project of 宗羽 (gxcsoccer). LoRA distillation (KL to teacher distributions plus cross-entropy), merged and quantised to 8-bit MLX, so it runs on Apple Silicon only; keep deskmind.json next to the weights. Option lists longer than 52 are scored in tournament rounds. In the product, brain-0.8b answers first and escalates about 70% of steps at threshold 0.96. The maker's own run of the third-party JevBench v1.4.2 public set gives 0.835 (easy 48/48, original 69/72, hard 76/111; the earlier G14 release scored 0.866) with Brier 0.269 and ECE 0.089; it reports no 13-gram overlap between its training mix and those items. Tuned for macOS Finder and TextEdit, web and form steps; the maker says the probabilities are the model's own weighting, not guarantees. The DeskMind name and logo are excluded from the licence. deskmind/eyes-4b is a GUI-grounding model, not a decision model.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 4B |
| Base model | Qwen/Qwen3.5-4B |
| Maker | DeskMind |
| Released | 2026-09-27 |
| License | apache-2.0 |
| Reported accuracy | 83.5% |
Get the weights
pip install systemonemodels
systemone pull deskmind/brain
The files are served from the maker's Hugging Face repository, deskmind/brain-4b, and verified against the checksums recorded here.