Moonshot made Kimi K3 available through its API on 16 July, and published the full weights on Hugging Face eleven days later, over 26 and 27 July. The repository is 118 files and roughly 1.56 terabytes. It is the first model in the three-trillion-parameter class whose weights anyone can download.
The published specification is unusually specific. K3 holds 2.8 trillion parameters in total but activates about 104 billion for each word it processes, drawing 16 experts from a pool of 896. Its context window is 1,048,576 tokens, and its attention stack combines 69 Kimi Delta Attention layers with 24 Gated Multi-head Latent Attention layers. On the model card it posts 93.5 on GPQA Diamond and 88.3 on Terminal-Bench 2.1, measured alongside Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5 and GLM-5.2.
It is worth being precise about where K3 actually lands, because the framing has been generous. On Moonshot's own fourteen-benchmark launch chart, K3 takes the top score on five rows and Claude Fable 5 takes seven. On FrontierSWE the two sit at 81.2 and 86.6. This is a model that is competitive with the closed frontier and still behind it on most measures. The story is the size of the gap, not the ranking.
The licence deserves a closer read than it will probably get. The weights ship under a bespoke document called the Kimi K3 License, tagged on Hugging Face as license:other rather than as a standard open-source licence. It is permissive for most people and most uses. Two conditions apply at scale: any product passing 100 million monthly active users or 20 million US dollars in monthly revenue must display the words Kimi K3 in its interface, and operators whose revenue exceeds 20 million US dollars over any consecutive twelve months must negotiate a separate agreement with Moonshot before commercial use. If you are shipping on top of K3, read the licence file itself rather than assuming the terms of the earlier K2 release carried over.
Reporting on the July wave of releases puts the capability gap between open and closed models at three to five months, down from a debated six to nine earlier in the year. That is the number worth carrying away, and it is a claim about pace rather than about any single model.
Our reading is that this changes less for individuals than the headlines suggest. A 1.56-terabyte download and the hardware needed to serve a 2.8-trillion-parameter model put self-hosting well out of reach of a laptop, and mostly out of reach of a single server. What has actually changed is the calculus for organisations that were already capable of hosting a large model and were choosing between an API bill and their own infrastructure. For them, the option on the table is now a genuinely frontier-adjacent one.
The most useful thing to understand from this release is not the 2.8 trillion figure but the 104 billion one. The gap between those two numbers is where nearly all of the past two years of efficiency work lives, and it is the reason a model can grow enormously without becoming proportionally more expensive to run.
Original source: Moonshot AI, via Hugging Face