Kimi-K3
A large language model developed by Moonshot AI that incorporates Mamba architecture components and is designed to handle long-context tasks.
Timeline (last 90 days)
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).
On the color prompt, all four Claude agents chose teal, while Kimi chose blue.
All players chose Samwise (a character from The Lord of the Rings) on the coffee companion prompt, indicating that models are difficult to distinguish on questions with high consensus.
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.
In the reverse experiment, four Kimi agents failed to identify the undercover Claude, and even labeled the real Claude as the least likely to be the mole.
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.
In the experimental setup, the four Claude agents had a maximum of 40 rounds of conversation to identify the infiltrating Kimi, while Kimi's task was to avoid being detected.
2 more timeline events
All Facts (12)
In 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%UnverifiedOn the color prompt, all four Claude agents chose teal, while Kimi chose blue.
50%UnverifiedAll players chose Samwise (a character from The Lord of the Rings) on the coffee companion prompt, indicating that models are difficult to distinguish on questions with high consensus.
50%UnverifiedIn the third set of tests, all four Claude agents chose 37 on the random number prompt, while Kimi chose 73.
50%UnverifiedKimi'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.
50%UnverifiedIn the experimental setup, the four Claude agents had a maximum of 40 rounds of conversation to identify the infiltrating Kimi, while Kimi's task was to avoid being detected.
50%UnverifiedFor Kimi-K3, the vLLM team compressed the latent tail down to the Mamba metadata level for performance optimization.
50%UnverifiedIn the reverse experiment, four Kimi agents failed to identify the undercover Claude, and even labeled the real Claude as the least likely to be the mole.
50%UnverifiedAn 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.
50%