In July last year, Moonshot AI directly open-sourced Kimi K2 — a trillion-parameter MoE large model that activates only 32B parameters per token.
Architecture Highlights
- Total parameters: 1 trillion
- Activation per token: 32B
- Training data: approximately 15.5 trillion tokens
- Context window: 128K (later upgraded to 256K)
- Optimizer: Muon
The code and weights were released under a Modified MIT License, making it one of the most permissive licenses among frontier-scale open-source models.
Performance Level
Among open-source non-reasoning (non-thinking) models, K2 ranks first on coding benchmarks and surpasses GPT-4.1 and Claude Opus 4 on multiple tasks. Its strengths lie in agent capabilities, tool calling, and STEM reasoning.
Two versions are available:
- K2-Base: for researchers who want to fine-tune it themselves
- K2-Instruct: for general conversation and agent scenarios, ready to use out of the box
Subsequent Iterations
In September, the K2-Instruct-0905 version was released, further improving coding task performance and expanding the context window from 128K to 256K.
In January this year, K2.5 was released — a continuous pre-training on the K2-Base foundation using approximately 15 trillion mixed visual and text tokens, natively supporting multimodality. It also includes a feature called Agent Swarm, which can decompose complex tasks into multiple parallel sub-tasks.
Judging by the iteration speed from K2 to K2.5, Moonshot AI is pursuing an aggressive open-source strategy. Open-sourcing a trillion-parameter model would have seemed like a fantasy two years ago.
Sources: Moonshot AI official technical blog, CocoLoop, Hugging Face model page