Kimi K3 Open Weights: License, Access, Reasoning Modes

Moonshot has published weights for each flagship in this line. Every generation uses a different license, and the K3 one is the most restrictive.

Hosted only means the Kimi app and kimi.com, Kimi Work, Kimi Code, and the Agent and Agent cluster modes. Moonshot lists K2.6, K3 and K3 Cluster as the app's model choices, with Agent modes layered on top for multi-step work. The release covers the model, the technical report and the training infrastructure. It does not hand you the hosted agent workflow, so running K3 yourself gives you the model, and you build the orchestration. Moonshot's own pages are where to check which product features have changed.

The K3 license is custom, so it is not interchangeable with MIT or Apache 2.0. The published text grants permission to use, copy, modify, merge, publish, distribute, sublicense and sell copies, and to run, deploy, fine-tune or create derivative works. Two conditions sit on top, and one exemption limits them:

• A Model as a Service business with revenue above 20 million US dollars (or the equivalent in other currencies) over any consecutive 12 months, counting the Licensee plus its affiliates, must enter a separate agreement with Moonshot AI before using the software or its derivative works for any commercial purpose.
• A commercial product or service with more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, must prominently display "Kimi K3" in its user interface.
• Section 4 exempts some use from those section 2 and 3 requirements: (a) internal use, meaning any use that does not make the software, its outputs, or its underlying capabilities available to third parties; or (b) any use of the software accessed through Moonshot AI's official products or certified inference partners.

The license's definition of Model as a Service does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.

K2 Thinking, K2.5, K2.6 and K2.7 Code use a Modified MIT License with the attribution condition only. The Model as a Service clause is new with K3. If you sell hosted inference at scale, that is the line to read closely, along with the exemptions and the definition above. If you fine-tune K3 for internal use or ship it inside a product below the attribution thresholds, nothing in the summary above should stop you, but check the exact wording against your case.

K2 Thinking was a separate model: you picked it by name, and the plain K2 models did not reason. K3 collapses that into one model. Per Moonshot's API docs, K3 always performs reasoning and returns it in reasoning_content, and preserved thinking is always on. K3 was trained in the preserved-thinking history mode, so follow Moonshot's guidance on carrying reasoning through multi-turn and tool-use conversations.

Depth is set with a top-level reasoning_effort field that accepts low, high or max. The default is max. Moonshot's docs say higher effort generally gives better answers and uses more tokens. K3 does not use the thinking parameter structure from earlier Kimi models, so code written for K2 Thinking needs changes.

The Kimi app exposes the same idea as a reasoning effort choice. Moonshot's help page tells you to pick Standard for speed and lower credit use, and Advanced or Extreme for harder analysis. The page does not say how those labels map to the three API values, so test both if you move between the app and the API. For what max effort costs in time and tokens, see the speed and verbosity measurements in the K3 review.

curl https://api.moonshot.ai/v1/chat/completions \
-H "Authorization: Bearer $MOONSHOT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
  "model": "kimi-k3",
  "reasoning_effort": "high",
  "messages": [{"role": "user", "content": "Your task here"}]
}'

The example uses the global endpoint, https://api.moonshot.ai/v1. The China-region endpoint is api.moonshot.cn, so use the one that matches where your API key was issued. The English docs are at platform.kimi.ai (platform.moonshot.ai now redirects there, and platform.kimi.com is Chinese by default). On the global platform, prices are in USD: $3.00 per 1M cache-miss input tokens, $0.30 per 1M cache-hit input tokens and $15.00 per 1M output tokens. The China platform prices the same tiers in CNY: 20, 2 and 100. The K3 review has the pricing comparison.

The weights are at moonshotai/Kimi-K3 on Hugging Face, and the code and license are in the MoonshotAI/Kimi-K3 repository on GitHub. Moonshot recommends vLLM, SGLang and TokenSpeed as inference engines and links recipes for each from the repository.

At 2.8 trillion total parameters, K3 is a multi-GPU, multi-node job. One community deployment report on the model's Hugging Face discussion board describes a K3 run on 16 NVIDIA H200 GPUs. That is a user report, not Moonshot guidance, so check the official recipes for supported configurations. NVIDIA has also published its own quantized checkpoint, nvidia/Kimi-K3-NVFP4. If you only want to try the model, the hosted API costs far less than a cluster. If your reason is control, data residency or fine-tuning, the weights are the point, and the license above is what to review first.

For smaller open-weight options you can run on modest hardware, see the open-weight model shortlist. For which model fits which task, use the task-by-task model guide.

Moonshot released Kimi K2 Thinking on November 6, 2025, as a 1 trillion parameter open-weight reasoning model with a 256K context window and native INT4 quantization. Its benchmark claims from that launch are out of date, so we removed them. The model is now on Moonshot's discontinued list alongside kimi-k2-thinking-turbo, and Moonshot recommends kimi-k3 in their place. Moonshot's wording for the original K2 preview and turbo models is "The kimi-k2 series models were officially discontinued on May 25, 2026." That does not cover the whole K2 family: kimi-k2.6 and kimi-k2.7-code are K2-family models and are not discontinued.

The weights are still on Hugging Face under the Modified MIT License. If you need an older, smaller-footprint Kimi you can self-host, the options are K2 Thinking, K2.5, K2.6 and K2.7 Code, all under the Modified MIT License. Of those, only K2.6 and K2.7 Code are still served by Moonshot's API. If you call Moonshot's API, move to kimi-k3 and replace any thinking parameter with reasoning_effort.

• Moonshot AI: Kimi K3 model card on Hugging Face
• Moonshot AI: Kimi K3 License
• Moonshot AI: Kimi K3 repository, recommended inference engines
• Moonshot AI: Kimi K3 weights, report, and infrastructure release
• Kimi API Platform: Kimi K3 quickstart, model id, reasoning_effort, reasoning_content
• Kimi API Platform: reasoning effort guide
• Kimi API Platform: thinking models guide, preserved thinking always on
• Kimi API Platform: model list, pricing and discontinued models
• Kimi Help Center: K2.6, K3, K3 Cluster, reasoning effort, and Agent
• Kimi Help Center: Agent Swarm
• Moonshot AI: Kimi K2 Thinking model card and Modified MIT License
• Moonshot AI: Kimi K2.6 model card
• Hugging Face discussion: community K3 deployment report on 16 H200 GPUs (user report, not Moonshot guidance)
• NVIDIA: Kimi-K3-NVFP4 checkpoint

Is Kimi K3 open source?

It is open weight. Moonshot released the K3 weights, technical report and training infrastructure on July 27, 2026, under the Kimi K3 License. That is a custom license, not a standard OSI license such as MIT or Apache 2.0.

What does the Kimi K3 License require?

It allows use, copying, modification, distribution, fine-tuning and commercial use. If you and your affiliates together run a Model as a Service business with more than 20 million US dollars in revenue over any consecutive 12 months (or the equivalent in other currencies), you must sign a separate agreement with Moonshot before using the software or its derivative works for any commercial purpose. That requirement does not apply to internal use, to any use that does not make the software, its outputs or its underlying capabilities available to third parties, or to use through Moonshot AI's official products or certified inference partners. The license definition of Model as a Service also excludes end-user products with model capabilities solely embedded within specific features or harnesses, and mere relaying of requests to models hosted by others. Commercial products with more than 100 million monthly active users or more than 20 million US dollars in monthly revenue must prominently display "Kimi K3" in the interface. This is a plain-language summary, not legal advice. Read the full LICENSE file.

Can you turn off thinking in Kimi K3?

No. Moonshot's API documentation says K3 always reasons. You control depth with the top-level reasoning_effort field, which accepts low, high or max, with max as the default. K3 does not use the thinking parameter that earlier Kimi models used.

Can I still use Kimi K2 Thinking?

Not through Moonshot's API. The kimi-k2-thinking model id is on Moonshot's discontinued list, and Moonshot recommends kimi-k3. The K2 Thinking weights remain on Hugging Face under a Modified MIT License, so you can still self-host them.

Written by

Paras Tiwari

Founder, Spectrum AI Labs

Founder of Spectrum AI Labs — testing AI tools and models, and writing up what actually ships.

More about Paras →