Kimi K3性能打平甚至超越Claude和Fable，价格却只有它们的五分之一，还开源，性价比炸裂。
Kimi K3是一款2.8万亿参数的模型，支持100万token上下文窗口，在SWE-Bench Pro、SWE-bench Multilingual、BrowseComp和Toolathlon上达到SOTA。它自主运行20小时内核优化实验，实现1.6倍加速。推理价格仅$0.25/百万token，而Fable 5收费$1.30。背后公司Moonshot估值$30B，计划于7月27日开源权重。
“The pricing moat that frontier labs have been sitting on just collapsed. You can now get frontier i...
“The pricing moat that frontier labs have been sitting on just collapsed. You can now get frontier intelligence at commodity prices, and it's only going one direction from here” – exactly what i predicted in August 2023. The tech community shunned me. Now you know why. Shruti @heyshrutimishra Two months ago, you paid a premium for Claude because nothing else came close. Then GPT-5.6 Sol matched Opus 4.8 on coding and agentic benchmarks. Yesterday, Kimi K3 beat them both. K3 is a 2.8 trillion parameter model with a 1 million token context window. It hits SOTA on SWE-Bench Pro, SWE-bench Multilingual, BrowseComp, and Toolathlon. It ran a 20-hour kernel optimization experiment with zero human intervention and achieved a 1.6x speedup on its own architecture. It costs $0.25 per million tokens. Fable 5 charges $1.30 for comparable performance. The company behind it, Moonshot, is valued at $30B. DeepSeek sits at $74B, GLM at $100B. Kimi built the best model at a fraction of its peers' valuations, while the US had them cut off from advanced chips. Open weights drop July 27. The pricing moat that frontier labs have been sitting on just collapsed. You can now get frontier intelligence at commodity prices, and it's only going one direction from here. 🔗 View Quoted Tweet 💬 8 🔄 18 ❤️ 81 👀 4492 📊 16 ⚡