Three weeks after Laya, Nandakishor M. has returned to the LocalLLaMA community with Vega, a physics-based typed decision model with 800 million parameters. Unlike a generative LLM, Vega does not produce text: it chooses between predefined outcomes using a physical metaphor of valleys, friction, and a rolling ball. The context window reaches 73,000 tokens, image support is included, and the author reports single-shot results that surpass many benchmarks from the earlier JEV architecture.

The internal mechanism is the most distinctive part. The full prompt—observation, question, possible outcomes such as YES/NO, and the last token—passes through an LLM/VLM. From its hidden layers, vectors representing these elements are extracted and projected to create a landscape with valleys, one for each outcome. A virtual ball is initialized in the valley, and friction slows its movement until it settles on the best outcome. Operationally, this means no fine-tuning for the task; instead, geometric representations of decisions are used at test time.

For self-hosted deployments, this reshapes some trade-offs. An 800-million-parameter model sits in a range that can run on consumer hardware without demanding the VRAM of much larger models. Because the decision happens at test time with no additional training, the pipeline is simpler: load the checkpoint, configure the engine, and classify over long contexts or images. Tasks such as detecting prompt injection attempts in system instructions, moderating content, or choosing between intents become natural candidates, since they require discrete and interpretable outcomes.

The structural signal is equally relevant. Vega emerged from the LocalLLaMA community, is open source, and moves decision-making into local infrastructure rather than behind a cloud API. If more low-latency decisions can be handled by small specialized models, the center of gravity shifts: data sovereignty and TCO reduction do not depend only on larger generative models, but also on lightweight decision layers that run close to the data. This direction rewards teams with modest hardware and strong control requirements, while eroding the advantage of centralized inference providers for routine classification work.

Of course, the benchmarks mentioned are tied to the JEV ecosystem and independent generalization still needs to be verified. But the combination of contained parameter count, long context, image support, and transparent decision mechanics is specific enough to avoid being confused with yet another generic clone. The project is documented on GitHub and Hugging Face, with a model card and a demo space for testing the Test-Time Training architecture.