Major open-weight model releases: Mistral Large 3, DeepSeek V4 Flash, Metis

Mistral Large 3 (675B MoE, Apache 2.0) is #2 OSS non-reasoning on LMArena, trained on 3000 H200s, multimodal/multilingual, optimized with NVIDIA/vLLM. DeepSeek V4 Flash delivers near-Claude Opus 4.8 coding performance at 99% cost reduction, accelerating the race to zero. Metis (4B/9B/27B open-source) introduces native memory as a first-class capability, challenging external memory pipelines. These releases underscore the commoditization of frontier capabilities and the shift toward open-weight models.