Kimi-K3
A large language model developed by Moonshot AI that incorporates Mamba architecture components and is designed to handle long-context tasks.
Core Facts
Timeline (last 90 days)
GPQA Diamond leaderboard top five: Darwin-180B-RSI 94.4% (180B), Kimi-K3 93.5% (2.8T), Darwin-397B-ZTC 93.4% (404B), Ornith-1.5-397B 92.8% (403B), Qwen3.8-2.4T-A95B 92.6% (2.4T).
In the command injection Java case, four of the six models passed the mount argument to ProcessBuilder as a list, while Qwen3.8-Max and Kimi-K3 built strings for sh -c, causing the shell to interpret user input as commands.
MLExpert's performance on MLE-Bench Lite approaches that of larger-scale frontier models such as Kimi-K3 and GLM-5.3.
MLExpert raised the base model's average medal rate on MLE-Bench Lite from 37.9% to 72.7%.
In LSCBench, Kimi-K3 had the highest functional pass rate (91.3%), but the lowest functional + security pass rate (16.4%).
NVIDIA Nemotron 3、Qwen3.5 至 Qwen3.8、Kimi-K3 以及 GLM-5.3-Flash 等模型均采用了混合架构范式。
Kimi-K3's pass@1 is 57.37% and its 20-of-20 is 17.60% (the Hugging Face blog post version reports pass@20 reaching 93.89% with a reliable portion of 13.41%).
Kimi-K3 achieved a pass@20 of 93.89% (476/507) on ThinkingBox, but its all-20 was only 13.41% (68/507).
An experiment was conducted on the NetMind Agent Arena platform in which four Claude Fable 5.1 agents and one disguised Kimi-K3 agent were placed in the same chat room.
Kimi's failure was due to a tendency to ask open-ended reflective questions, combined with a failure to consistently use the commit-reveal sealed mechanism, which allowed Claude to answer later and blend into the group.
6 more timeline events
All Facts (9)
MLExpert raised the base model's average medal rate on MLE-Bench Lite from 37.9% to 72.7%.
75%UnverifiedGPQA Diamond leaderboard top five: Darwin-180B-RSI 94.4% (180B), Kimi-K3 93.5% (2.8T), Darwin-397B-ZTC 93.4% (404B), Ornith-1.5-397B 92.8% (403B), Qwen3.8-2.4T-A95B 92.6% (2.4T).
50%UnverifiedIn the command injection Java case, four of the six models passed the mount argument to ProcessBuilder as a list, while Qwen3.8-Max and Kimi-K3 built strings for sh -c, causing the shell to interpret user input as commands.
50%UnverifiedMLExpert's performance on MLE-Bench Lite approaches that of larger-scale frontier models such as Kimi-K3 and GLM-5.3.
50%UnverifiedIn LSCBench, Kimi-K3 had the highest functional pass rate (91.3%), but the lowest functional + security pass rate (16.4%).
50%UnverifiedNVIDIA Nemotron 3、Qwen3.5 至 Qwen3.8、Kimi-K3 以及 GLM-5.3-Flash 等模型均采用了混合架构范式。
50%UnverifiedKimi-K3's pass@1 is 57.37% and its 20-of-20 is 17.60% (the Hugging Face blog post version reports pass@20 reaching 93.89% with a reliable portion of 13.41%).
50%UnverifiedKimi-K3 achieved a pass@20 of 93.89% (476/507) on ThinkingBox, but its all-20 was only 13.41% (68/507).
50%UnverifiedFor Kimi-K3, the vLLM team compressed the latent tail down to the Mamba metadata level for performance optimization.
50%