The big move this week was making NEURA's brain model agnostic. Up to now it leaned on one model and that was that. If the model was slow, NEURA was slow. If it was down, NEURA was down. That never sat right with me for something you are supposed to own, so I pulled the model out of the core and put a provider layer in its place.
What that means in practice is NEURA can now reach for the best cloud models when you want real power behind a hard question, and it can run a fast private model locally through Ollama when you would rather keep everything on your own machine. Same companion, same voice, but you get to pick the engine underneath it.
The piece I am most proud of is the tiered router. The idea is simple: send the hard problems to a deep model that can actually think, send the quick back and forth to a fast one so it feels snappy, and keep a fallback ready so NEURA is never stuck staring at a dead connection. Most turns do not need a heavyweight model, and it felt wasteful to treat them like they do.
The tricky part was the seams. Every provider talks a little differently, and getting them to look identical to the rest of NEURA took more plumbing than I expected. But it was worth it.
Bring your own brain is core to what makes this different. NEURA is not one locked-in assistant you rent. It is yours, and now the mind inside it is yours to choose too.