Vega open-sources a physics-based typed decision model: frozen Qwen3.5-0.8B plus a 57 MB engine that rolls a ball into calibrated answers

Decision models became a category this week (OpenAI Decisions API beta, Microsoft Decision-1, TypeSafe AI's Jev raise). On Oct 10 indie developer Nandakishor M open-sourced Vega (Apache 2.0, pip install vegaml), following his earlier Laya release.

Architecture: a frozen Qwen3.5-0.8B (or 4B), pinned by SHA-256, is read at layers 13 and 19; the layers above and the LM head never run, so no text is generated. One prompt holds the observation, the question and the candidate answers, and five pooled span vectors feed a 57 MB trained engine. That engine builds one potential-energy valley per answer and rolls a damped ball for 12 steps in 64 dimensions; where it settles is the answer, with a calibrated probability, a conformal set and an abstain flag. It supports a 73,728-token context with prefix caching, plus image input.

Author-reported numbers: typed-decisions test 0.763 (0.8B + adapter) and 0.803 (4B) at 22/28 ms; phishing 75.4 acc / 252 of 400 caught vs Jev 1.13.0 at 61.9 / 99; median latency 267 ms on a T4 vs 591 ms hosted. The author notes Jev leads on most other tasks, including reranking and multi-step reasoning, and Vega scores only 0.240 on MMMU-Pro.

Use it for local classification and routing where data cannot leave the machine; fine-tune adapters via the included Kaggle 2xT4 notebook.