They seem to have held back the “big” locally runnable model.
It’s also kinda conservative/old, architecture wise: 16-bit weights, sliding window attention interleaved with global attention. No MTP, no QAT (yet), no tightly integrated vision, no hybrid mamba like Qwen/Deepseek, nothing weird like that. It’s especially glaring since we know Google is using an exotic architecture for Gemini, and has basically infinite resources for experimentation.
They seem to have held back the “big” locally runnable model.
It’s also kinda conservative/old, architecture wise: 16-bit weights, sliding window attention interleaved with global attention. No MTP, no QAT (yet), no tightly integrated vision, no hybrid mamba like Qwen/Deepseek, nothing weird like that. It’s especially glaring since we know Google is using an exotic architecture for Gemini, and has basically infinite resources for experimentation.
It also feels kinda “deep fried” like GPT-OSS to me, see: https://github.com/ikawrakow/ik_llama.cpp/issues/1572
IMO it’s not very interesting, especially with so many other models that run really well on desktops.