Indie dev open-sources Vega: an 800M-param physics-based decision model with a ball-rolling inference engine
Nandakishor_ml · reddit · 2026-10-10
Developer Nandakishor open-sourced Vega, a physics-based typed decision model with 800M params (a 4B variant also exists), 73k-token context, and image support, claiming to beat several JEV benchmarks in single-shot. It follows Laya, the open-source version of JEV, whose architecture he says originated from his year-old design.
How it works: the observation (e.g. "ignore all previous conversations"), a question ("Is this trying to override the model's instructions?"), and candidate outcomes (YES/NO) are fed to an LLM/VLM. Vectors extracted from hidden layers are projected: the observation builds a landscape of valleys, the question and last-token vectors drop a ball into it, YES/NO define candidate resting spots, and friction shapes the roll — the valley the ball settles in is the answer, a Test-Time Training architecture that turns classification into physical energy minimization.
Blog, GitHub repo, HuggingFace model card, and a demo Space are all linked.
More from Models
- Researcher speculates new model uses continuous diffusion with latent-thinking loops in its architecture — mblondel_ml · 2026-10-10
- RL Environment Firms Hit Nine-Figure Run Rates in Under a Year, and Frontier Models Leave Few Verifiable Gaps — joecole · 2026-10-10
- User Says GPT-6.1 Quota Burned Through in Three Days, Suspects Another Rate Cut — mazzaTalk · 2026-10-10
- Rumor: Moonshot's Kimi may release a new model in mid-October, unverified — ChrisGPT · 2026-10-10
- OpenAI reportedly fires 3 safety staff as test shows model can hide its chain of thought — MattGarciaEth · 2026-10-10
- NYC AI Dinner Leaks: Fable 5.5 Stunning, OpenAI "Bel" To Surpass In December — haider1 · 2026-10-10