Open-weight model competition intensifies

Chinese models (Kimi K3, Qwen3.8 Max, DeepSeek V4 Flash) and Korean 700B models (A.X K2, K-EXAONE 2.0) are closing the gap with frontier labs. Muse Spark 1.2 (Meta) beats GPT-5.5 on math. Cost advantages of 2-10x drive market shift. Open-weight releases from Western labs (gpt-oss, Nemotron) add to the landscape. New: Kimi K3 biology assessment shows it's strongest open-weights biology model but still 8 months behind closed frontier and refuses only 26.9% of hazardous prompts. A practical guide for selecting open-weight models for RL post-training was published. U.S. DOE launches Genesis Open Models initiative, signaling government backing for open-weight science models. Alibaba announces revenue-sharing model for Qwen3.8, creating three-tier licensing landscape. A new article highlights that Chinese open models dominate Hugging Face downloads (41%) and provided security advantage when commercial models refused to analyze exploit logs. Today: A Kimi K3 article from The Hindu describes it as a landmark release triggering US soul-searching and an open letter signed by major players, while Anthropic refuses. Also: Jevons Paradox article argues cheap open-weight models expand total demand. MiniMax H3 video generation model released. DeepSeek V4 Flash 0731 achieves 82.7% on Terminal-Bench 2.1. Qwen3.8-27B confirmed for next week.