China Open-Sourced a 2.8-Trillion-Parameter Model — And It Lands Three Points Behind the Best AI on Earth

The largest AI model anyone has ever released to the public is not from OpenAI, Anthropic, Google or Meta. It is 2.8 trillion parameters, it weighs more than a terabyte, it was built in Beijing, and as of late July you can download the whole thing.

Moonshot AI unveiled Kimi K3 in mid-July and put the weights on Hugging Face days later, just ahead of the July 27 date it had promised. The files run to 96 shards and well over a terabyte. Nobody is running this on a gaming PC.

The specification is the headline, but it is not the story. The story is where K3 lands on the scoreboard, and what the licence attached to it actually says.

What Moonshot Actually Shipped

K3 is a mixture-of-experts model: 2.8 trillion total parameters with 104 billion activated per token. Mixture-of-experts is why that first number is less alarming than it sounds — the model holds an enormous pool of specialised sub-networks and routes each token to a small fraction of them, so the compute cost per token tracks the 104 billion, not the 2.8 trillion.

It takes a 1,048,576-token context window — a hair over a million tokens, enough to hold a large codebase or a shelf of documents in a single conversation. It is natively multimodal, accepting text, images and video rather than bolting vision on afterwards. And it always reasons, with the reasoning budget now exposed as low, high or max rather than the single maximum setting its predecessor offered.

That API price is worth sitting with. Three dollars in, fifteen out, is exactly Claude Sonnet 5's standard rate. A Chinese lab has stopped pricing to undercut and started pricing to match.

Three Points From the Top

On the Artificial Analysis Intelligence Index, K3 scores 57. That is the highest result any open-weight model has posted — the next one down, GLM-5.2, sits at 51. Claude Fable 5 scores 60.

Three points. That is the entire gap between the best model on Earth that you have to rent, and the best model on Earth that you can download.

On some things the gap has already closed the wrong way. K3 beats Fable 5 outright on the Frontend Code Arena leaderboard with 1,679 Elo. On the BenchLM composite it ranks 5th of 214 models at 79.9 out of 100, and its strongest category is multimodal and grounded work — screenshots, documents, charts — where it places 2nd of 33.

Fifth overall is not first, and Moonshot has not claimed otherwise. K3 still trails Anthropic's Claude and OpenAI's GPT on aggregate performance. But fifth is a different conversation from the one the industry was having a year ago, when the honest summary of Chinese open models was "impressive for the price."

The $20 Million Clause

Kimi K2 shipped in July 2025 under a modified MIT licence. K3 did not. It ships under a custom document called the Kimi K3 Licence, and the difference matters commercially.

The licence permits downloading, self-hosting, fine-tuning and quantisation. But operators running the model as a service must sign a separate agreement with Moonshot before commercial use if their group revenue exceeds $20 million over any consecutive twelve months.

For a solo developer, a research lab or a mid-sized company running K3 internally, that threshold is invisible. For a cloud provider that wants to resell K3 inference, it is a negotiation with a Beijing company you cannot avoid. Which is, presumably, the point.

Compare it to Google's Gemma 4, released in April under Apache 2.0 with no revenue ceiling at all. Both are "open." Only one of them is open in the sense that a lawyer would sign off on without a phone call.

Then Congress Called DoorDash

The predictable thing happened. A model this good, this cheap, and this downloadable gets used — and on July 31 that collided with Washington.

DoorDash had disclosed that it routes harder engineering tasks to a US-developed frontier model while assigning lower-level work to Kimi K2.6, K3's predecessor. The company said the split configuration outperformed its earlier setup while costing less. That is a straightforward engineering decision, and a lot of companies are quietly making it.

John Moolenaar, chairman of the House Select Committee on China, and Andrew Garbarino, chairman of the House Committee on Homeland Security, wrote to DoorDash chief executive Tony Xu asking for a full accounting of every Chinese model the company uses and every security test run against them. Documents are due August 14. In-person briefings from the staff responsible for AI infrastructure, software security, model evaluation, procurement and compliance are due August 21.

The lawmakers pointed to public statements from the White House Office of Science and Technology Policy alleging that Moonshot operated a covert platform to conduct large-scale distillation against American AI models, and trained on advanced computing systems it was not authorised to obtain. Moonshot has not publicly answered those specific claims.

That sentence is the whole problem in one clause. The lawmakers are not arguing that the engineering logic is wrong. They are conceding it is right, and asking what happens next.

The Money Behind It

Moonshot is not a scrappy outsider any more. The company reportedly hit a $20 billion valuation in May on a $2 billion raise, and is said to be raising another $1 billion to $2 billion at a valuation of up to $31.5 billion — roughly 50% higher, in about three months. Bloomberg reported on July 22 that Moonshot is planning a final pre-IPO round at a pre-money valuation of up to $50 billion, with talks starting this month and a Hong Kong listing possible within six months.

Those numbers describe a company that intends to be permanent. And the release cadence backs it up: K2 in July 2025, K2.5 in January, K2.6 in April with a vision encoder and the ability to fan a complex task out across as many as 300 parallel agents, K3 in July. The original K2 series was formally discontinued on May 25 — Moonshot is now deprecating its own models fast enough that keeping up is a job.

What to Take From This

• The open-weight gap is now three points, not three generations. Whatever premium you are paying for a closed frontier model, you should be able to say what it buys you. A year ago that was an easy question.
• Read the licence before you build on it. "Open weights" is a marketing category, not a legal one. The $20 million service-provider threshold in the K3 licence is invisible until suddenly it isn't.
• Nobody can un-release this. Export controls govern chips. They do not govern a torrent of weights that thousands of people have already downloaded. Once the files are public, the policy question stops being whether to allow it and becomes how to live with it.
• Mixed-model routing is now a compliance question. DoorDash made a defensible cost decision and got two committee chairmen and a deadline. If you are routing any workload to a Chinese open-weight model, the time to document your security testing is before somebody asks.
• Watch pricing, not benchmarks. K3 matching Sonnet 5's headline rate instead of undercutting it is the single most confident thing Moonshot did this month.