Open-weight efficiency, specialized models, and typed inference reshape price-performance

DeepSeek, Qwen, GLM, Gemma, Kimi, Mistral, local inference, low-bit serving, and specialized models challenge the assumption that every workflow needs a large frontier model. Aleph Alpha's Kolibri adds a European sovereign/open-weight signal with 78B total and roughly 3.5B active parameters, but reported tool-use and coding weaknesses require independent validation. Recent llama.cpp, Diffusers, and Ai2 OlmoCore 3 results show practical infrastructure gains, while capability, energy, licensing, and deployment advantages remain incompletely validated.