Model Positioning
Kimi-K2-Base is a pretrained base language model in the Kimi K2 series, primarily intended for researchers and developers who want to fine-tune it themselves and build custom solutions. It differs from the post-trained Kimi-K2-Instruct version, which is intended for general chat and agent interactions. (kimi.com)
Architecture and Capabilities
- Uses a mixture-of-experts (MoE) architecture with approximately 1 trillion total parameters, of which approximately 32 billion are activated per forward pass; the official model card lists a context length of 128K tokens. (huggingface.co)
- The official source states that pretraining used approximately 15.5 trillion tokens. Official evaluations of the Base version cover general knowledge, programming, mathematics, and Chinese-language tasks; for example, MMLU is 87.8 (5-shot) and LiveCodeBench v6 is 26.3 (1-shot). These are Base version results and should not be confused with Instruct version scores. (github.com)
Uses and Availability
Suitable for base model research, continued training, fine-tuning, and developing custom language model applications. The model weights have been officially released, and the code and weights are available under a modified MIT license; users should read the full license and its additional conditions before use. Officially listed inference engines include vLLM, SGLang, KTransformers, and TensorRT-LLM. (github.com)
Limitations
Base is a base model, not the Instruct version optimized for direct chat and tool calling; practical applications typically require task-specific adaptation and evaluation. The original Kimi K2 announcement stated that vision capabilities were not yet supported at the time and noted that difficult reasoning or unclear tool definitions could lead to excessive token generation, truncated output, or incomplete tool calls; these are series-level statements and should not be treated as separate empirical results for the Base version. (kimi.com)
Pricing
As of 2026-11-18, the official domestic pricing page and official international pricing page were checked, but no official API input or output prices explicitly applicable to Kimi-K2-Base were verified, so no amounts are listed; prices for other Kimi K2 versions cannot be directly treated as Base prices. Self-deployment may also incur computing and storage costs. (platform.kimi.com)
Release Date and Official Sources
- Release date: 2025-07-11; commit history in the official model repository can be traced back to that date. (huggingface.co)
- Official Kimi K2 announcement
- Official Kimi-K2-Base model card and weights
- Official code repository and license