Kimi K2

Kimi K2 is a family of large language models developed by Moonshot AI, introduced in July 2025. Its defining features are a mixture-of-experts architecture, approximately one trillion total parameters, and about 32 billion parameters activated for each token. The family was developed for general language tasks, programming, and AI agents that interact with external tools. Moonshot released both pretrained and instruction-following model weights. (arxiv.org)

Architecture

Kimi K2 uses a sparse Transformer architecture. Rather than applying every parameter to every input, its mixture-of-experts layers route each token through a selected subset of expert networks. The technical report gives approximately 1.04 trillion total parameters, conventionally rounded to one trillion in model summaries, with roughly 32 billion active per token. This distinction separates the model’s total capacity from the computation used for an individual token. (raw.githubusercontent.com)

The original model’s published specifications include the following: (github.com)

Multi-head latent attention compresses the representations retained for attention, reducing the memory associated with cached keys and values during inference. Sparse expert activation reduces per-token computation relative to activating the entire network, but does not eliminate the need to store or distribute its much larger collection of weights. Consequently, active-parameter counts alone do not describe deployment requirements. (raw.githubusercontent.com)

Training

The initial K2 model was pretrained on 15.5 trillion tokens. A token is a unit produced by tokenization, rather than necessarily a complete word. Its training introduced MuonClip, an optimization method combining the Muon optimizer with a stabilization mechanism called QK-Clip. QK-Clip addresses excessive query–key attention scores, which can destabilize large-scale training. The developers reported completing pretraining without loss spikes. (arxiv.org)

Post-training emphasized the transition from generating responses to carrying out tasks. The training process included large-scale synthesis of tool-use interactions and reinforcement learning involving both real and simulated environments. These interactions supplied examples of selecting tools, interpreting their outputs, and continuing toward a task objective. The technical report describes a joint reinforcement-learning stage intended to improve both general responses and agentic behavior. (arxiv.org)

Principal variants

Kimi-K2-Base is the pretrained checkpoint, intended as a foundation for research, fine-tuning, and customized systems. Kimi-K2-Instruct is the post-trained checkpoint for conversational assistance and tool use. The original Instruct release was designed to respond without a dedicated extended-thinking phase; this distinguishes its operating style from the later Thinking variant, rather than implying an absence of reasoning capability. (github.com)

Kimi-K2-Instruct-0905, announced in September 2025, extended the context window from 128K to 256K tokens. Its model card emphasizes improvements in agentic coding and frontend development. The developers reported 69.2% on SWE-bench Verified, compared with 65.8% for the initial Instruct checkpoint, under the evaluation conditions described in that release. (forum.moonshot.ai)

Kimi K2 Thinking adds extended reasoning interleaved with tool calls and retains a 256K-token context window. Its post-training incorporates quantization-aware training for native INT4 deployment. Moonshot reported experiments involving 200–300 sequential tool calls, presenting the variant as a model for longer research, programming, and writing workflows. These are reported capabilities under particular test conditions, not guaranteed outcomes for every task. (huggingface.co)

Tool use and applications

K2’s tool-calling interface allows an application to provide descriptions and parameter schemas for available functions. The model produces a structured request; the surrounding software executes the function and returns its result to the conversation. Thus, tool use depends on an external execution system: the model does not acquire unrestricted access to files, networks, or programs merely by generating a call. (github.com)

Documented applications include repository-level software engineering, code generation, command-line tasks, and multi-step information gathering. The distinction between the model and its execution framework is important: prompts, tools, permissions, and the number of permitted attempts form part of the evaluated system. (arxiv.org)

Evaluation and limitations

For the initial Instruct model, Moonshot reported 65.8% on SWE-bench Verified in a single-attempt agentic setting, 53.7% on LiveCodeBench v6, and 75.1% on GPQA-Diamond. The original evaluation primarily compared models without extended thinking. These figures describe specific benchmark versions and protocols, rather than a universal ranking of model quality. (arxiv.org)

Later releases used different reasoning budgets and tool configurations. For example, K2 Thinking’s published results distinguish tool-free reasoning from tool-assisted evaluation and higher-computation settings. Its model card also notes that the consumer chat interface used fewer tools and shorter tool-use sequences than some benchmark configurations, so chat behavior might not reproduce the reported scores. (huggingface.co)

Deployment precision is another source of variation. Moonshot’s vendor-verification project compares tool-call responses from external providers against its own service and identifies model versions, inference-engine versions, and tool-call formatting as relevant compatibility factors. A shared model name therefore does not by itself establish identical implementation or behavior across providers. (github.com)

Distribution and licensing

The initial checkpoints were distributed in block-FP8 format, with deployment documentation covering inference frameworks including vLLM, SGLang, KTransformers, and TensorRT-LLM. This permits independent deployment rather than restricting use to a hosted application programming interface. (huggingface.co)

The code and weights use a Modified MIT License. In addition to the underlying permission and notice provisions, the license requires prominent “Kimi K2” branding in commercial products or services exceeding either 100 million monthly active users or US$20 million in monthly revenue. Accordingly, the release should not be described as using the unmodified MIT License; “open-weight” identifies the availability of the parameters without implying that its terms are identical to standard open-source software licenses. (github.com)

References

1. Kimi K2: Open Agentic Intelligencearxiv.org
2. GitHub - MoonshotAI/Kimi-K2github.com
3. Kimi K2 — Technical Reportraw.githubusercontent.com
4. Kimi K2-0905 updateforum.moonshot.ai
5. moonshotai/Kimi-K2-Instruct-0905huggingface.co
6. moonshotai/Kimi-K2-Thinkinghuggingface.co
7. GitHub - MoonshotAI/K2-Vendor-Verifiergithub.com
8. Kimi-K2/LICENSEgithub.com