For most of the past three years, the frontier AI conversation has been an American one: OpenAI, Anthropic, Google. That changed quickly in July 2026, when a Beijing company called Moonshot AI released Kimi K3 — a 2.8-trillion-parameter model that briefly topped a major public coding leaderboard and then published its own weights for anyone to download. Kimi AI is now the most serious open challenger to the closed Western labs, and it is available free at kimi.com.
Here is what Kimi actually is, which of its models are worth using, what it genuinely does better than ChatGPT and Claude — and the places it still loses.
What is Kimi AI?
Kimi is an AI chatbot and family of large language models built by Moonshot AI, a Chinese company founded in 2023. You use it like ChatGPT — through a web app at kimi.com, a mobile app, or an API. Its defining trait is that Moonshot publishes the model weights, so anyone can download and run them.
That last point is the whole story. OpenAI and Anthropic ship their best models only as a service you rent. Moonshot ships its best model as a file. According to public records of the company’s releases, it has done this consistently since the first open-weights Kimi K2 in July 2025.
Which Kimi model is the best right now?
Kimi K3 is the best Kimi model as of July 2026. It has 2.8 trillion total parameters, a 1-million-token context window, and native vision. For cheaper everyday work, K2.6 costs roughly five times less per output token, and K2.7 Code is the specialised coding variant.
The current line-up breaks down like this:
- Kimi K3 — released 16 July 2026, weights published 27 July 2026. A sparse Mixture-of-Experts model with 896 experts, 16 of which fire on any given token. Context window is exactly 1,048,576 tokens. This is the flagship, and it powers chat, agent mode, and Kimi Code.
- Kimi K2.7 Code — a June 2026 coding-focused variant. Moonshot reported a 21.8% improvement over K2.6 on its own Kimi Code Bench v2 while using about 30% fewer reasoning tokens per agentic loop.
- Kimi K2.6 — the April 2026 workhorse. A 1-trillion-parameter MoE with 32 billion parameters active per token, and the first Kimi model to accept video input. Still the value pick at $0.95 / $4 per million input/output tokens.
The short version: use K3 when the task is hard, K2.6 when the task is high-volume, and K2.7 Code when you are writing software.
What does Kimi AI do better than ChatGPT and Claude?
Kimi’s advantages are open weights, price, and agents. It is the only frontier-class model you can legally download and self-host, its API costs roughly 40% less than comparable Claude and GPT tiers, and its OK Computer agent mode ships a full virtual computer — browser, terminal, file system — inside the chat window.
1. You can actually download it
The K3 weights live at huggingface.co/moonshotai/Kimi-K3 — roughly 594 GB in native MXFP4 format. For a company with data-residency rules or a genuine air-gap requirement, that is the difference between “we cannot use frontier AI” and “we can”. No US lab offers this at the top of its range.
Read the licence before you plan a business around it, though. It is not plain MIT. Moonshot published a bespoke “Kimi K3 License” that requires a separate negotiated agreement if you resell K3 as a model-as-a-service above a $20 million annual revenue threshold, and it requires the words “Kimi K3” to appear in your product’s interface past certain thresholds.
2. It undercuts the competition on price
K3’s API runs $3 per million input tokens and $15 per million output tokens, with cache hits dropping to $0.30. That sits below the comparable Claude Opus and GPT-5.6 Sol tiers, which are reported around $5 in / $25–30 out. On the consumer side, the free tier gives unlimited basic chat — a meaningfully more generous starting point than most rivals.
3. OK Computer is a real agent, not a demo
Kimi’s agent mode, launched in late 2025, gives the model its own virtual machine and more than 20 tools: a file system, a browser, a terminal, a code interpreter, plus image and audio generation. From one prompt it will build multi-page websites, produce editable slide decks, and turn hundreds of thousands of rows of spreadsheet data into interactive visualisations.
K2.6 added an “Agent Swarm” layer on top, scaling to 300 domain-specialised sub-agents that coordinate across as many as 4,000 steps in a single autonomous run. If you have compared assistants for real work before — as we did in ChatGPT Work vs Claude Cowork — this is the category Kimi is competing in, not simple Q&A.
4. The benchmarks are not marketing fluff
K3 ranked first out of 99 models on WebDev Arena with an Elo of about 1,678 — the first open-weight model ever to top that leaderboard. Independent evaluator Artificial Analysis placed it 4th of 189 models on its Intelligence Index, and 2nd overall on private evaluations with an Elo of 1,547.
Where Kimi AI still falls behind
Moonshot is unusually honest about this: the company’s own materials place K3 behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on overall performance, while beating Claude Opus 4.8 and GPT-5.5 across coding and agentic tests. It is a strong second tier, not the outright leader.
There are practical rough edges too. Developer Simon Willison, testing K3 on release day, noted that the model ships with only one reasoning effort level — maximum — so it burns tokens even on simple prompts. A single trivial test generated over 13,000 reasoning tokens and cost 25 cents. There is also no free API tier; the developer API needs a minimum $1 top-up to activate.
Self-hosting is realistic only for well-funded teams. Moonshot recommends a supernode of 64 or more accelerators. “Open weights” does not mean “runs on your laptop”.
Finally, there is the question every buyer outside China asks: data governance. Using kimi.com means sending your prompts to a Chinese company’s servers. Self-hosting the weights sidesteps that entirely — which is precisely why the open release matters more than the benchmark scores.
Who should actually use Kimi AI?
Three groups get clear value. Developers running high-volume API workloads will save real money on K2.6 and K3. Teams with data-residency or air-gap requirements finally have a frontier-class option they can host. And anyone doing long-document work — legal review, codebase analysis, research synthesis — benefits from the 1-million-token context window.
If you are a casual user who just wants a chatbot that works, ChatGPT and Claude remain smoother products with better ecosystems. Kimi’s free tier is worth ten minutes of your time, but it is not yet a reason to cancel anything. The same “try before you switch” logic applies to any fast-moving tool category — see our breakdown of Veo vs Gemini Omni vs Google Vids.
Frequently Asked Questions
Is Kimi AI free to use?
Yes. The free plan at kimi.com includes unlimited basic chat plus a small allowance of agent credits, but excludes Agent Swarm and Kimi Code. Paid memberships start at $19 per month, with higher tiers at $39, $99 and $199, all around 20% cheaper on annual billing.
Is Kimi K3 better than ChatGPT?
Not overall. Moonshot’s own benchmarks put K3 ahead of GPT-5.5 on coding and agentic tasks but behind GPT-5.6 Sol on general performance. Where K3 clearly wins is price and the fact that you can download and self-host it.
Is Kimi AI open source?
It is open-weight rather than fully open-source. You can download the K3 weights from Hugging Face, but the bespoke Kimi K3 License adds a revenue-triggered clause for model-as-a-service resellers and a user-interface attribution requirement.
Can I run Kimi K3 on my own hardware?
Only with serious infrastructure. The weights are about 594 GB in MXFP4 format and up to 1.4 TB depending on quantisation, and Moonshot recommends 64 or more accelerators. Smaller Kimi models are far more practical for local deployment.
How much does the Kimi API cost?
Kimi K3 is $3 per million input tokens and $15 per million output tokens, dropping to $0.30 for cache hits. Kimi K2.6 is significantly cheaper at $0.95 / $4. There is no free API tier — activation requires a minimum $1 recharge.
