Moonshot AI Releases Kimi K3 Weights: 2.8 Trillion Parameters, Free to Download

Moonshot AI has published the complete model weights for Kimi K3, putting a 2.8-trillion-parameter system into the hands of anyone willing to download it. The files appeared on Hugging Face on the evening of July 26, arriving slightly early against the company's own July 27 commitment. That closes an eleven-day window that began with K3's API debut on July 16.

The scale is the headline. K3 now ranks as the largest open-weight model in the world that anyone can grab for free, fine-tune, and run on their own hardware. Moonshot has described it as the first open-weight system to press up against the 3-trillion-parameter class, built with advanced reasoning, long-horizon coding, and knowledge work as its target uses.

Inside the Architecture: 1M Context and Sparse Activation

The raw parameter count only tells part of the story. K3 runs on a Mixture-of-Experts design, meaning it doesn't fire all 2.8 trillion parameters on every token. Roughly 50 billion parameters activate per token instead, which keeps inference costs in a range that makes the model practical rather than theoretical.

Paired with that is a 1-million-token context window. For the long-horizon coding tasks Moonshot points to, that window is arguably as important as the parameter count. It lets the model hold entire codebases, extended research materials, or multi-stage task chains in working memory without the constant re-summarizing that shorter windows force.

The combination — enormous total capacity, sparse activation, and a very long context — is what distinguishes K3 from earlier open releases that competed on size alone.

Demand Outran Moonshot's Own GPUs

The API launch on July 16 made the infrastructure problem obvious fast. Inside 48 hours, usage was pushing against the ceiling of Moonshot's compute clusters. The company responded by pausing new consumer subscriptions entirely.

In a post on X, Moonshot framed it plainly: the reception had gone well beyond what the team expected, and the GPUs were under real strain. The New York Post reported the announcement, which landed on July 17 during Shanghai's World Artificial Intelligence Conference. The reasoning behind the pause was protective rather than punitive — existing paying customers kept their service quality while the company worked to add capacity behind the scenes.

That's an unusual position for an AI lab to be in publicly. Most capacity crunches get managed quietly through rate limits. Shutting the door on new signups signals that the constraint was hard, not cosmetic.

Together AI and Modal Went Live on Day Zero

Cloud providers didn't wait. Together AI and Modal both timed hosted access to coincide with the weight drop, meaning developers had managed endpoints available the moment the files went public.

That day-zero availability matters more than it might seem. Open weights are only useful to teams that can actually run them, and for most organizations the practical path to a 2.8-trillion-parameter model runs through someone else's infrastructure.

What It Actually Takes to Self-Host K3

Downloading the model is free. Running it is not.

In its native MXFP4 format, K3 comes in at roughly 594 GB. That figure sets the floor for what self-hosting demands:

Four H100s or comparable hardware is a serious capital commitment, not a weekend project. The open-weight label is genuine, but it describes freedom to use and modify rather than freedom from infrastructure cost. Small teams and individual developers will mostly encounter K3 through hosted providers.

Benchmark Results Place K3 Near the Frontier

Independent evaluation from Artificial Analysis put K3's intelligence score just below Anthropic's Fable 5 and above OpenAI's GPT-5.5. On the Frontend Code Arena leaderboard, K3 took the top position outright.

Those results are the substance behind the attention. An open model landing between two leading closed systems on general intelligence scoring — and beating both in a specialized coding arena — narrows the performance argument that closed labs have relied on.

Pricing Undercuts Closed Rivals by Roughly Half

The commercial angle sharpens the picture further. K3's API pricing sits at $3 per million input tokens and $15 per million output tokens. The Associated Press, citing Bank of America analysts, noted that this comes to roughly half the cost of OpenAI's GPT-5.6 Sol.

Half the price at comparable capability is a difficult combination for competitors to answer. And with the weights now free to download, buyers gain a third option beyond paying Moonshot's API rate or a rival's premium: running it themselves and paying only for compute.

The Open-Weights Escalation and a Hong Kong IPO

Reuters has reported that Moonshot has opened conversations with Goldman Sachs and other institutions about a potential IPO in Hong Kong. The weight release lands in the middle of that process, functioning as both a technical event and a demonstration of market position.

Analyst Nathan Lambert has given the broader pattern a name: the open-weights escalation. His framing describes a dynamic where open models keep climbing in capability and, in doing so, squeeze the pricing power and strategic moat that closed-source competitors have depended on. Each release that closes the performance gap makes proprietary access harder to charge a premium for.

K3 is a clear data point in that direction. A model this large, this capable, and this freely available shifts the negotiating position of every buyer evaluating AI infrastructure — and it does so at a moment when the competition between Chinese and American labs is being watched closely on both technical and financial terms.