Maincode releases Matilda Decision v1, a 27B multimodal decision model
Apache-2.0 weights on Qwen3.8-27B. The maker's own run scores 65.38 on the public Jev Decision Index 0.3 suite.
Maincode published Matilda Decision v1 on Hugging Face on 15 October 2026, by repository creation date. The card describes a 27B multimodal decision model that takes text or JSON, optionally with images, and answers several questions in one call. It returns a probability for each answer and generates no text. Question types are multiple choice up to 255 options, yes/no and rating.
It is post-trained from Qwen3.8-27B with supervised fine-tuning, then reinforcement learning. About 26.1B parameters, BF16, Apache-2.0. Loading needs transformers 5.17.0 and trust_remote_code.
The maker ran the Jev Decision Index 0.3 public suite itself: 65.38 overall across 140,178 requests. The card's table puts Torchcast Decision 27B at 65.10, Perplexity Decider v1.1 at 62.25 and Jev at 57.96.
The limits matter. The model must be served with bidirectional attention; a causal engine gives different and worse answers. Requests over 131,072 tokens raise an error. Images are capped at about 1.6 MP. No inference provider deploys it today.
The lead over the next model is 0.28 points on a self-run index. Treat the ranking as a draw.
Sources
- https://huggingface.co/Maincode/matilda-decision-v1
- https://huggingface.co/spaces/multimodalart/jev-decision-index
Editor's first pass (AI-assisted)
- Real 3.5: Weights are public; serving needs bidirectional attention and custom code, and no provider hosts it.
- Useful 3.5: Routing, yes/no and urgency rating from text, JSON and images.
- Notable 3: Edges a rival on the board by a fraction of a point.
- Novel 3: Another entry in the same format.
Comments