Kimi K2
The previous-generation large language model released by Moonshot AI, featuring a total of 1 trillion parameters and a Mixture-of-Experts architecture. It held the top position among open-source models by parameter count for an extended period.
Timeline (last 90 days)
Models tested and subsequently eliminated include Mistral 24B, GPT-OSS 20B, Kimi K2, Phi-4, DeepSeek, and Cydonia, mostly due to being too slow, performing poorly, or having better alternatives.
SWE-2 is a coding model built by Cognition on top of Moonshot AI's Kimi K2 model, released on September 10, with a primary focus on cost efficiency
Kimi K2 is an open-source Mixture-of-Experts model released by Moonshot AI in 2025, with approximately 32 billion activated parameters and more than 1 trillion total parameters
The research team compared seven BPE-based tokenizers within a single language: GPT-2, GPT-4, gpt-oss, Llama 3.1, Gemma 3, Qwen3, and Kimi K2.
Models such as Kimi K2 have already exceeded 2TB in scale.
OpenAI used Semi-Analysis's public benchmark InferenceX to test Jalapeño on three open models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2 1T.
The UP host stated that they had used Kimi K2 and Alipay-related model plans and did not encounter any instances of quota burning at extreme speed.
In recent years, the pace of open-sourcing domestic large models has noticeably accelerated, from Alibaba's Qwen series to DeepSeek and then to Moonshot's Kimi K2.
All Facts (8)
Kimi K2在FP16精度下仅模型权重就需要数百GB显存，INT4量化可将显存需求压缩至约四分之一
85%UnverifiedModels tested and subsequently eliminated include Mistral 24B, GPT-OSS 20B, Kimi K2, Phi-4, DeepSeek, and Cydonia, mostly due to being too slow, performing poorly, or having better alternatives.
50%UnverifiedKimi K2 is an open-source Mixture-of-Experts model released by Moonshot AI in 2025, with approximately 32 billion activated parameters and more than 1 trillion total parameters
50%UnverifiedThe research team compared seven BPE-based tokenizers within a single language: GPT-2, GPT-4, gpt-oss, Llama 3.1, Gemma 3, Qwen3, and Kimi K2.
50%UnverifiedModels such as Kimi K2 have already exceeded 2TB in scale.
50%UnverifiedOpenAI used Semi-Analysis's public benchmark InferenceX to test Jalapeño on three open models: GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2 1T.
50%UnverifiedIn recent years, the pace of open-sourcing domestic large models has noticeably accelerated, from Alibaba's Qwen series to DeepSeek and then to Moonshot's Kimi K2.
50%UnverifiedSWE-2 is a coding model built by Cognition on top of Moonshot AI's Kimi K2 model, released on September 10, with a primary focus on cost efficiency
50%