Moonshot AI released its flagship language model Kimi K3 on 16 July 2026. The official site of the product is kimi.ai for international users and kimi.com for the Chinese market. The model weights are open and sit on Hugging Face, and the license permits commercial use if you meet its conditions. Below: the specifications, how to get access, and a comparison with Claude by type of task, with no invented benchmarks.
What Kimi K3 is
Moonshot AI calls K3 "the first open 3T-class model" (official blog; the entry in the Kimi publications index is dated 16 July 2026). The architecture is a mixture of experts with 896 experts; for each token 16 of them switch on, plus two shared ones. The model has 2.8 trillion parameters in total, and 104 billion work on every inference step (huggingface.co/moonshotai/Kimi-K3). The context window reaches 1,048,576 tokens. The model is natively multimodal and takes text, images and video.
The same blog says Kimi Delta Attention and Attention Residuals were added to the architecture on top of Stable LatentMoE. By the authors' estimate that gives about a 2.5x gain in scaling efficiency over the previous generation, Kimi K2.
Open weights and license
Moonshot AI promised to publish the full weights by 27 July 2026, and they are now available on Hugging Face, with the code on GitHub. The license is called the Kimi K3 License. It allows you to use, copy, modify and distribute the model, commercially as well. There are two thresholds, and past them you need a separate agreement with Moonshot AI. The first: a business's total revenue on this model above $20 million over any 12 consecutive months under the Model as a Service scheme. The second: products with an audience above 100 million MAU or revenue above $20 million a month must show the name "Kimi K3" in their interface.
The K2 family (Kimi-K2-Instruct, K2.5, K2.6, K2.7-Code) ships under a similar Modified MIT License with the same thresholds.
Kimi K2.8 Preview: what this release is
On 11 September 2026 reports appeared of a full launch of Kimi K2.8 Preview in Kimi Code and Kimi Work. The model is presented as an intermediate step between K2.7 Code and the coming K3 class (according to Pandaily, 11 September 2026). We found no official post about this release on moonshot.ai at the time of writing, so treat the K2.8 details as preliminary.
Kimi Slides: presentations from text
Kimi Slides is part of the Kimi+ set next to Docs, Sheets, Websites and Code (see kimi.ai/kimiplus). The scheme is simple: describe the content in words, attach a style sample if you like, and the service builds the slides. It all opens right on kimi.ai or kimi.com with nothing to install, and basic generation costs nothing beyond the subscription's usual credit pool.
There is also an agent mode for multi-page sites and presentations, which SCMP wrote about. Which model actually powers slide generation is not stated officially anywhere. Third-party reviews name Kimi K2 Thinking, but the official site does not confirm it, so that is a guess, not a fact.
Kimi K3 and Claude: which model for which tasks
We found no direct official "Kimi K3 vs Claude Opus 5" benchmark from either Moonshot AI or Anthropic, and we will not make up figures. The open specifications and the license terms do allow some practical conclusions, though.
Kimi K3 makes sense when you need open weights: you can download the model, deploy it yourself and fine-tune it for your task, and in most commercial scenarios the license does not forbid that. For research teams and for products where control over the infrastructure matters, that is a real advantage.
Claude stays a closed model, available only through the Anthropic API or chat. But in agentic tasks, where the model decides for itself what to do next (write code, call tools, work in a terminal), the ecosystem around Claude is noticeably more mature today: Claude Code, MCP, broad IDE integration. If that is your case, comparing the models on parameters alone is pointless. What has been built around them matters more.
Kimi in codexkey
codexkey has Kimi. The catalog as of 29 October 2026 lists kimi-k3 ($0.62 / $3.08 per 1M tokens, input / output, 1M context), kimi-k3.1-pro ($0.42 / $2.06, 1M context) and kimi-k2.7-code ($0.2 / $0.8, 262,144 context). The Kimi platform says that from 31 October 2026 the K2 series will be unavailable to new users of its API, and that all K2 models are planned to be switched off by the end of December 2026 (platform.kimi.ai). We do not know what will happen to kimi-k2.7-code in the codexkey catalog, so for a new project it is wiser to take K3 or K3.1 Pro.
You can connect Kimi through Claude Code and codexkey: payment from Russia without a VPN, and you switch the model with the command /model kimi-k3. Setup goes through environment variables:
export ANTHROPIC_BASE_URL="https://api.codexkey.online"
export ANTHROPIC_AUTH_TOKEN="sk_cdx_YOUR_KEY"
export ANTHROPIC_DEFAULT_OPUS_MODEL="claude-opus-5-5[1m]"
export ANTHROPIC_DEFAULT_SONNET_MODEL="claude-sonnet-5[1m]"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="claude-sonnet-5"
claude
The ANTHROPIC_DEFAULT_* lines are required. Without them Claude Code asks for its own default models and answers "There's an issue with the selected model".
Step-by-step instructions for Linux, macOS and Windows are in the Claude Code guide. For the Kimi Code tool itself, how to install it and how it differs from Claude Code, read Kimi Code: installation and use. And if what you care about is paying for Kimi from Russia, go to Kimi subscription and payment from Russia.
FAQ
Where is the official Kimi site?
The international version is kimi.ai. The Chinese one is kimi.com. They are two domains of one Moonshot AI product: kimi.ai prices subscriptions in dollars, kimi.com in yuan. Documentation and the API platform live at platform.kimi.ai.
Which country is Kimi from?
China: Kimi is made by Moonshot AI. The company has a corporate site, moonshot.ai, and a Chinese-language version, moonshot.cn, while the product for ordinary users runs on the separate domains kimi.ai and kimi.com.
What is Kimi Slides, and do I pay for it separately?
Kimi Slides generates presentations from a text description and is part of the Kimi+ set. For basic generation there is no charge beyond the usual Kimi subscription: the tool spends the account's shared credit pool.
Did Kimi K3 beat Claude or GPT in tests?
We found no official comparative benchmarks from Moonshot AI against specific Anthropic or OpenAI models. The claims about Kimi K3's place in rankings that circulate online come from unofficial aggregators, not from the Moonshot blog. We do not repeat them here.
Can I download and run Kimi K3 myself?
Yes. The weights are open and sit on Hugging Face, with the code on GitHub. The Kimi K3 License permits commercial use, but a large business (revenue above $20 million over 12 months on the model, or an audience above 100 million MAU) needs a separate agreement with Moonshot AI.
Is Kimi ai free?
Basic chat on the K2.6 model is available without a subscription and does not spend paid credits. Agentic tasks, Deep Research and coding are limited by a quota (according to user reports; the site has no official description of the free plan's limits). K3 and the advanced modes need a paid subscription.