What it is
Kimi K3 is Moonshot AI's flagship model, launched in the Kimi app and API on 16 July 2026, with open weights on Hugging Face from 27 July. Moonshot calls it the world's first open 3-trillion-class model: a mixture-of-experts design with 2.8 trillion parameters, 104 billion active per token, trained with 4-bit (MXFP4) weights from fine-tuning onward. It reads text, images and video, writes text, and has a 1,048,576-token context.
Thinking is always on, with low, high or max effort (max by default). The weights use a modified MIT licence: free for most uses, but companies running a model-as-a-service business with over $20 million in revenue over 12 months need a separate agreement, and products with over 100 million monthly users or $20 million in monthly revenue must display the Kimi K3 name.
On Moonshot's API (platform.kimi.ai) it costs $3 per million input tokens, $0.30 for cached input, and $15 per million output tokens. Amazon Bedrock, Together and Fireworks also serve it.
Our review
Kimi K3 is the strongest open-weight model for long agent work, and among the most expensive to use. Moonshot's card compares it at max effort with Claude Fable 5, GPT-5.6 Sol and Claude Opus 4.8, the leading closed models of July 2026. It led on BrowseComp (91.2), MCPMark-Verified (94.5) and SWE-Marathon (42.0), and came within a point of Fable 5 on OSWorld-Verified. It trailed Fable 5 on Humanity's Last Exam (43.5 against 53.3) and FrontierSWE (81.2 against 86.6), and GPT-5.6 Sol on DeepSWE (67.5 against 73.0). These are Moonshot's runs, some on internal tests.
Independent measurement narrows the lead. Artificial Analysis scores it 44 on its Intelligence Index, fifth in its open-weight class, at $2.00 per index task and 37 output tokens a second on Kimi's API, 72nd of 78 for speed. GLM-5.3-Flash scores 42 at $0.28 per task. In Surge AI's blind coding evaluation, published by Mistral, K3 scored 3.59 out of 5, just behind GLM-5.3 and Mistral Large 4.
The data terms deserve a careful read. Moonshot's privacy policy lets it use user content, including prompts and outputs, to train its models, describes no API exemption or opt-out, and keeps data while the account is active on servers in Singapore. Teams with sensitive inputs should use the weights themselves or a host such as Amazon Bedrock, which lists the same $3/$15 price.
Self-hosting is a cluster project. The checkpoint is about 1.56 TB, and the vLLM recipe sets a floor of eight Nvidia GB300 or AMD MI355X GPUs, with multi-node layouts for real traffic. Integrators also need to return the full assistant message, reasoning included, on every turn, because K3 was trained with preserved thinking history and Moonshot says multi-turn and tool calls require it.
Where it earns its place
A research-tools startup building a deep-research agent
Answer questions that need dozens of web searches and long reports
K3's 91.2 on BrowseComp and 1M context suit long browse-and-summarise loops; Moonshot's own run used context compaction at 300K tokens. At $3/$15, a task with 500K input and 30K output tokens costs about $1.95 before caching.
A fintech platform team that cannot send customer data to Moonshot
Use a frontier-class open model under its own cloud contract
Call K3 on Amazon Bedrock at the same $3/$15, with a 128K output cap there, instead of platform.kimi.ai, whose policy permits training on user content.
An AI lab or large enterprise with a GPU cluster
Self-host an open model that competes with closed frontier models on agent tasks
Serve the 1.56 TB MXFP4 checkpoint with vLLM or SGLang on at least eight GB300 or MI355X GPUs, multi-node for production. Internal use is exempt from the licence's revenue and display clauses.
Before you commit
- Setup
- Select kimi-k3 on platform.kimi.ai through its OpenAI- or Anthropic-compatible API, or use Kimi Code CLI, which Moonshot recommends as the agent harness. Pass the full assistant message, including reasoning_content and tool_calls, back on each turn.
- Integrations
- Vision input, structured output, tool choice, dynamic tool loading and context caching on Moonshot's API; also on Amazon Bedrock, Together, Fireworks and others via OpenRouter. vLLM, SGLang and TokenSpeed for self-hosting.
- Data and privacy
- Moonshot's privacy policy allows training on user content with no API exemption described, keeps data while the account is active and stores it in Singapore; no SOC 2 or ISO certification is named. The licence exempts internal use from its revenue and display clauses.
- Cost at scale
- $3/$15 per million tokens, $0.30 cached, with extra charges for 5-minute or 1-hour cache writes. 1,000 agent tasks of 500K input and 30K output tokens cost about $1,950 uncached; Fireworks lists one endpoint at $2.10/$10.50.
- Limits
- 1,048,576-token context; most OpenRouter hosts allow 943K output tokens, Bedrock 128K. Thinking cannot be disabled, and Artificial Analysis measured 37 tokens a second with a 3.5-second first token on Kimi's API.
What's good
- Leads Moonshot's table on BrowseComp, MCPMark and SWE-Marathon
- Open weights; licence conditions only bite above $20M revenue
- 1M-token context with image and video input
- Same $3/$15 on Moonshot, Bedrock and Together
What's not
- $2.00 per Artificial Analysis task, near the bottom of its class
- Slow at 37 tokens a second on Kimi's API
- Moonshot's policy allows training on API content, with no stated opt-out
- Self-hosting needs at least eight GB300 or MI355X GPUs
Best for
Teams building browsing and research agents who want open weights and can self-host or pay frontier-like API prices.
Alternatives
Kimi K3: common questions
- Is Kimi K3 worth it?
- Kimi K3 tops Moonshot's own agent scorecard, but at $2 per Artificial Analysis task it costs seven times GLM-5.3-Flash for two more index points. We rate it 4.3 out of 5.
- How much does Kimi K3 cost?
- Open weights under a modified MIT licence. Kimi API: $3 input / $15 output per 1M tokens; cached input $0.30.
- What are the downsides of Kimi K3?
- $2.00 per Artificial Analysis task, near the bottom of its class; slow at 37 tokens a second on Kimi's API; moonshot's policy allows training on API content, with no stated opt-out; self-hosting needs at least eight GB300 or MI355X GPUs.
- Who is Kimi K3 best for?
- Teams building browsing and research agents who want open weights and can self-host or pay frontier-like API prices.
- How much does Kimi K3 cost for a team or in production?
- $3/$15 per million tokens, $0.30 cached, with extra charges for 5-minute or 1-hour cache writes. 1,000 agent tasks of 500K input and 30K output tokens cost about $1,950 uncached; Fireworks lists one endpoint at $2.10/$10.50.
- Is Kimi K3 safe for company data?
- Moonshot's privacy policy allows training on user content with no API exemption described, keeps data while the account is active and stores it in Singapore; no SOC 2 or ISO certification is named. The licence exempts internal use from its revenue and display clauses.
- What are the best Kimi K3 alternatives?
- The closest alternatives we have reviewed are GLM-5.3, Qwen3.8-Max, DeepSeek-V4-Pro and MiniMax-M3.
Answers checked Oct 23, 2026.
Make Kimi K3? Show this rating on your site
