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

Kimi K2

Open mixture-of-experts model for agentic tasks and coding

Moonshot AI released Kimi K2 in July 2025 as a large language model with one trillion total parameters. It uses thirty-two billion active parameters per token and focuses on autonomous tool use and software engineering.

Share on WhatsApp

  • One trillion total parameters with 32B active
  • Trained on 15.5 trillion tokens using MuonClip
  • 65.8% pass rate on SWE-bench Verified
  • Released under Modified MIT License
  • Optimized for agentic tool use and coding

What it is

Kimi K2 is a large language model built by Moonshot AI. It arrived in July 2025 with a total size of one trillion parameters. The design uses a mixture of experts approach, meaning only thirty-two billion parameters activate for each token. This structure aims to balance high capacity with efficient computation. The model targets complex reasoning, coding, and autonomous agent tasks where it must use external tools. Developers can download the weights or access the system through an application programming interface.

How it came about

The Kimi Team at Moonshot AI developed the model over several years. They published a technical paper in July 2025 detailing its construction. The team focused on scaling up training without the instability often seen in large models. They released two main variants: a base version for researchers to fine-tune and an instruct version for general chat and agent work. Updates in late 2025 improved the chat template and tokenizer robustness.

Inside the model

The architecture relies on sixty-one layers with multi-head latent attention. It selects eight experts from a pool of three hundred and eighty-four for each token. Training used fifteen point five trillion tokens and a custom method called MuonClip. This optimizer helps prevent loss spikes during the learning process. Post-training involved reinforcement learning and synthetic data generation to improve how the model interacts with real environments and executes code.

What it can do

The model performs strongly in software engineering benchmarks. It scores sixty-five point eight percent on SWE-bench Verified tests using standard tools. In multilingual coding tasks, it reaches forty-seven point three percent. Mathematical reasoning results include forty-nine point five on AIME 2025 and seventy-five point one on GPQA-Diamond. These scores apply to settings without extended thinking time, positioning it among top open-source models for agentic work.

How to use it

Moonshot AI provides the model weights under a Modified MIT License. Users can find the checkpoints on Hugging Face in block-fp8 format. The company also offers an API compatible with OpenAI and Anthropic standards. Recommended inference engines include vLLM and SGLang. The instruct variant suggests a temperature setting of zero point six for optimal performance in chat applications.

Questions people ask about Kimi K2

What is the total size of Kimi K2?

The model has one trillion total parameters. However, it uses a mixture of experts design, so only thirty-two billion parameters are active for any given token. This allows for high capacity while maintaining computational efficiency during inference.

Can I use Kimi K2 for commercial projects?

Yes, the model weights and code are released under the Modified MIT License. This permissive license generally allows for commercial use, modification, and distribution, though users should review the specific license text for any conditions.

How does Kimi K2 handle coding tasks?

It achieves strong results on coding benchmarks like SWE-bench. The model is designed for agentic workflows, meaning it can autonomously decide when to call tools, such as code editors or bash terminals, to solve complex software engineering problems.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Kimi K2 yet, so the facts on this page come from the sources listed above: huggingface.co/moonshotai/Kimi-K2-Instruct, arxiv.org/abs/2507.20534. 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.