China’s Kimi K3 Just Took the Top Coding Spot, and You Can Download It

For a couple of years now, the story about Chinese AI has been that it is close, impressive, and always a step or two behind the American frontier. That story just got harder to tell. Moonshot AI, the Beijing startup behind the Kimi chatbot, has released Kimi K3, and it did not quietly match the leaders. It walked into the Frontend Code Arena, a public leaderboard where models compete head to head on building working interfaces, and finished first. Ahead of Claude. Ahead of GPT. At the very top.

Here is the part that should make you sit up: this is not a locked API you rent from a company an ocean away. Kimi K3 is open-weight, which means the full model is going out to anyone who wants it. Moonshot says the complete weights land on July 27, and at 2.8 trillion parameters, K3 is the largest openly released model the world has seen. You could, in theory, download the same thing that just beat everyone at coding and run it yourself.

What it actually is

Kimi K3 is a mixture-of-experts model, which is a fancy way of saying it is enormous on paper but only fires a small slice of itself for any given request. Of its 896 internal experts, it activates just 16 per token, so you get the knowledge of a 2.8-trillion-parameter brain without paying to run all of it at once. It also handles up to a million tokens of context (roughly a very long book's worth of text at a time) and can see, with native vision built in rather than bolted on.

The number people keep pointing at is 1,679. That is K3's score in the Frontend Code Arena, enough for the number one slot and enough to edge out Claude Fable 5, which had been sitting comfortably at the top. On coding and agentic tasks more broadly, Moonshot's model beat Claude Opus 4.8 and GPT 5.5. For a sense of how fast this moved: the previous Kimi, K2.6, ranked 18th. K3 jumped 17 places in a single generation.

The honest ledger

Now the caveats, because there are real ones. Winning one leaderboard is not the same as being the best model overall, and K3 is not. On the broader private evaluations that try to measure general capability, it still trails Claude Fable 5 and OpenAI's GPT 5.6 Sol. Being first at frontend coding is a genuine achievement and also a specific one. Read the headline for what it says, not for what it implies.

There is a cost wrinkle too. K3 has a single reasoning effort level, and it likes to think. In one early hands-on test, a simple prompt burned through more than 13,000 reasoning tokens and cost about a quarter just to answer. API pricing sits at $3 per million input tokens and $15 per million output tokens, which is right in line with the mid-tier Western models rather than the bargain that earlier Kimi releases were. Cheap to download is not the same as cheap to run.

Why this one matters

Strip away the benchmark horse race and the interesting thing is where this came from. Moonshot says K3 was trained partly on export-grade Nvidia H200 chips and partly on hardware from an unnamed alternative vendor, and that it squeezed roughly 2.5 times more scaling efficiency out of that setup than its predecessor thanks to a new attention design. In plain terms: faced with U.S. limits on the best chips, a Chinese lab built a top-tier model anyway, then gave it away.

That is the real signal here, and it is aimed at two different audiences. If you build with AI, an open model that codes at the frontier is a gift, because it means you are no longer renting your best option from a handful of American companies that can change the terms whenever they like. If you run one of those companies, K3 is a warning shot. The gap that everyone assumed would protect the frontier just got a lot thinner, and the model that closed it is one anybody can download.