Kimi新开源模型K3，参数2.8万亿，能处理百万上下文，速度还快了6倍多，做长程智能体任务很合适。
Kimi发布K3开源模型，拥有2.8万亿参数、100万上下文窗口和原生多模态能力。K3采用Kimi Delta Attention，在百万token上下文中实现解码速度提升6.3倍。Attention Residuals机制以不到2%的额外成本带来约25%的训练效率提升。该模型面向长程智能体编程和自演进工作流设计，开源权重计划于2026年7月27日开放。
Bullish on open models. At this rate, it won't even make sense to compare models directly. And it...
Bullish on open models. At this rate, it won't even make sense to compare models directly. And it looks like I was right. We might have an open-weight Fable/Mythos 5-level model by EOY. But don't get distracted because overdependence on one model is a poor strategy. Combine models instead. How I decide which models to use now is based on what they can solve and how effective and efficient they are. No single benchmark can measure that. Thus, private evals on your harnesses and workflows are the only real way to tell how much you are leaving on the table when you exclude models in the top 10 in this chart. The difference in results that matters doesn't change that much once you start fusing model capabilities. Very few people know this. I say this because you would be absolutely mistaken if you were only relying on one or the top model in this chart. Capability gaps between models have been significantly shrinking, and that trend will accelerate in the coming months. Kimi.ai @Kimi_Moonshot Introducing Kimi K3: Open Frontier Intelligence 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost 🔹 Built for long-horizon agentic coding and self-evolving workflows Kimi K3 is now live on Kimi.com xZUo, Kimi Work, Kimi Code, and the Kimi API. Open Weights by July 27, 2026. 🔗 platform.kimi.ai XAqMw 🔗 Tech kimi.com/blog/kimi-k3 MSNM1f 🔗 View Quoted Tweet 💬 5 🔄 2 ❤️ 10 👀 1514 📊 5 ⚡