Kimi K3: a Chinese lab gave away its best AI model. Here's what that means

There are two kinds of AI model. Most of the famous ones — ChatGPT, Claude, Gemini — are closed: you can use them through an app or an API, but you never get the model itself. A smaller group are open-weight: the company publishes the actual model file so anyone can download it, examine it, adapt it, and run it on their own machines. On 27 July 2026, the Chinese lab Moonshot AI published the biggest open-weight model anyone has ever released.

Kimi K3 has 2.8 trillion parameters — the internal values a model learns during training. For scale, that is roughly 75% larger than DeepSeek's V4 Pro, which was itself considered enormous. It can hold a little over a million tokens of context at once, which is about 750,000 words: a long novel, or an entire codebase, in a single conversation.

A clever design choice keeps it usable. Rather than running all 2.8 trillion parameters for every word, K3 splits itself into 896 specialist sections and activates only sixteen of them at a time — about 104 billion parameters per token. Think of it as a large hospital where you see two or three relevant specialists rather than every doctor in the building.

In parts, genuinely yes. In a blind evaluation where developers compared outputs without knowing which model produced them, K3 came first for building website front-ends — ahead of Anthropic's Claude Fable 5, one of the most expensive models on the market. On a business-tasks benchmark called AA-Briefcase it came second, beating OpenAI's GPT-5.6 Sol Max.

But Moonshot also published a comparison table showing its model losing to Claude Fable 5 on a number of tests, including software engineering and difficult reasoning. That is unusually honest of them, and it is the more accurate summary: K3 is excellent at some things and a clear step behind on the hardest ones.

Not on your laptop. Moonshot recommends at least 64 specialist AI chips to run K3, and the model file is around 1.4 terabytes. That puts self-hosting out of reach for individuals and most companies. What you can do is use it through a hosted service, where it costs roughly $3 per million words of input and $15 per million of output — about a fifth of what Claude Fable 5.1 charges.

One correction worth knowing, because it circulated widely and was wrong: K3 is not released under the permissive MIT licence. It uses a custom Kimi K3 License with its own terms, which matters if you plan to build a business on it.

The bigger picture is about price. When a free-to-download model can do most of what an expensive one does, the expensive one has to justify its premium on the hardest work only. Days after K3, Alibaba announced it would openly release a 2.4-trillion-parameter model of its own — a size it had always kept private before. For anyone buying AI, that competition is straightforwardly good news.

There are a couple of rough edges worth knowing about before trusting it with important work. Moonshot says the model can lose quality when the software running it trims its internal reasoning to save space — a common technique in tools that chain many steps together. It also launched with only one 'thinking effort' setting, permanently set to maximum, which means even simple requests do the full amount of work and cost accordingly. One developer found a single small image-generation test cost 25 cents. Cheaper settings were promised in later updates.

None of that changes the headline: a model this capable being published openly by a Chinese lab, and then matched days later by a rival Chinese lab doing the same, is the clearest sign yet that the advantage held by the big American AI companies is narrowing rather than widening.