Moonshot AI releases Kimi K2
Moonshot AI's trillion-parameter MoE model, continuing its emphasis on lossless recall across a very long context window.
CURATED TIMELINE · EDITORIAL EDITION
Follow the market’s shifts from bilingual chat models and hyperscaler-backed open families to sparse MoE efficiency, controllable reasoning, and long-context products across ChatGLM, Qwen, DeepSeek, and Kimi.
Timeline overview
Editorial thread
Reviewed event briefs and original editorial context, ordered to show how the story changed over time.
Moonshot AI's trillion-parameter MoE model, continuing its emphasis on lossless recall across a very long context window.
Qwen3 combines fast direct answers and longer reasoning in one open-weight family, spanning dense and MoE models up to a 235B flagship.
DeepSeek publishes R1, R1-Zero, six distilled models, and its reinforcement-learning recipe, pushing open reasoning models into the frontier conversation.
DeepSeek releases a 671B MoE model with 37B active parameters, open weights, low API prices, and unusually detailed efficiency disclosures.
Qwen2.5 spans small local models through a 72B flagship, with stronger instruction following and specialist code and math editions.
DeepSeek releases a 236B mixture-of-experts model that combines sparse activation with latent attention to cut training and inference costs.
ChatGLM3 adds native function calling and code execution to a compact 6B model, shifting the ChatGLM line from conversation toward agent-like workflows.
Alibaba Cloud releases Qwen as a multilingual open-weight foundation-model family, giving developers a major China-built alternative across model sizes.