News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnKimi K2 Thinking

Kimi K2 Thinking

Open-source reasoning model with tool use and INT4 quantization

Moonshot AI released Kimi K2 Thinking as an open-source mixture-of-experts model designed for deep reasoning. It uses native INT4 quantization to lower latency and supports long sequences of tool calls for autonomous tasks.

Share on WhatsApp

  • Uses mixture of experts with 32B active parameters
  • Supports 256k token context window
  • Native INT4 quantization for lower latency
  • Stable tool use for up to 300 steps
  • Open weights available for download

What it is

Kimi K2 Thinking is a large language model built by Moonshot AI. It functions as a thinking agent that reasons through problems step by step. The system can call external tools dynamically during this process. It is designed to handle complex workflows such as coding and research. The model uses a mixture of experts architecture. It activates 32 billion parameters per token from a larger pool. This design helps manage computational costs while maintaining high performance. The model supports a context window of 256,000 tokens. It is available with open weights for the community.

How it came about

This model follows the earlier K2 0905 version. Moonshot AI developed it to improve multi-step reasoning depth. The team focused on creating a system that remains stable over long interactions. Previous models often lost coherence after thirty to fifty steps. K2 Thinking aims to sustain goal-directed behavior for up to three hundred consecutive tool invocations. It represents the latest iteration in the Kimi series of thinking models. The release emphasizes open access to the model weights.

Inside the model

The architecture uses a mixture of experts approach. It contains sixty-one layers in total. Only one layer is dense, while the rest use sparse experts. There are 384 experts available, with eight selected for each token. One expert is shared across all inputs. The attention mechanism is MLA, and it uses SwiGLU activation. A key feature is native INT4 quantization. The developers applied Quantization-Aware Training during the post-training stage. This method reduces inference latency and GPU memory usage without significant performance loss. The model can hide previous tool outputs if the context limit is exceeded.

What it can do

The model performs well on reasoning benchmarks. On Humanity's Last Exam with tools, it scored 44.9. It achieved 99.1 on AIME25 when using Python. In heavy mode, it reached perfect scores on AIME25 and HMMT25. It also scored 60.2 on BrowseComp with tools. For coding tasks, it reached 71.3 on SWE-bench Verified. The heavy mode uses parallel trajectories to aggregate results. This allows it to match or exceed other leading models in specific tests. It maintains stability during long sequences of tool calls.

How to use it

Moonshot AI provides the model weights openly. Users can download them from platforms like Hugging Face. The model supports native INT4 quantization for efficient deployment. This allows for faster inference on standard hardware. The company notes that chat modes on their website may use fewer tools. This means web-based interactions might not reflect full benchmark capabilities. An agentic mode is planned to expose the full feature set.

Questions people ask about Kimi K2 Thinking

What is the parameter size of Kimi K2 Thinking?

The model activates 32 billion parameters per token. It uses a mixture of experts architecture with 384 total experts. Eight experts are selected for each token during processing.

How does the model handle long contexts?

It supports a context window of 256,000 tokens. If tool outputs exceed this limit, the system employs a strategy to hide previous outputs. This helps maintain coherence during long tasks.

Is the model available for local use?

Yes, the model weights are open. Users can download them from repositories like Hugging Face. It supports INT4 quantization to reduce memory requirements for local deployment.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Kimi K2 Thinking yet, so the facts on this page come from the sources listed above: huggingface.co/moonshotai/Kimi-K2-Thinking. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.