Vega: An 800M Parameter Physics-Based Typed Decision Model Released as Open Source

Developer Nandakishor_ml has open-sourced Vega, an 800-million-parameter physics-based typed decision model (with a 4B parameter variant available) supporting multimodal image input and a 73k token context window. Built under the Laya project, Vega utilizes a Test-Time Training (TTT) engine and adapter architecture to map inputs directly into physics-based decision landscapes. By substituting traditional autoregressive generation with a vector-based physics simulation, Vega provides a lightweight open-source alternative to proprietary decision models like Jev. Its compact size and long context capability enable efficient, low-latency, and non-hallucinating classification tasks on local hardware. Vega extracts observation, question, outcome, and token vectors from internal LLM/VLM hidden layers to build a simulated landscape with outcome valleys (such as YES or NO). Decision selection is modeled as a ball rolling down this energy landscape with friction until it comes to rest in the winning candidate outcome valley.