Model updates

Official news, first rumor reports and substantive progress. Each entry keeps the information published at that time.

• vegaml open-sources Vega, an 800M/4B physics-based typed decision model with 73k context and images

  NandhaKishorM released vegaml on GitHub and PyPI: typed decisions from a frozen language-model backbone plus a small physics engine, returning choice/score/boolean answers with calibrated probabilities, conformal sets, and abstain—without generating text. Defaults are an 800M (Qwen3.5-0.8B frozen) build with a 14.3M engine; a 4B option is in the same checkpoint repo. Context is 73,728 tokens with read-once prefix caching, and an image path is supported. The README cites LocalLLaMA/typed-decisions test-split figures (for example soft accuracy 0.680 on 0.8B and 0.725 on 4B) and says inference on a single T4 is demonstrated in Colab.

  OfficialOfficial updateOct 10Source: GitHub @NandhaKishorM/vegamlPart of Vega

RSS for this view →