Moonshot's 2.8 trillion parameter Kimi K3 open weights land tonight at 1.4TB

Moonshot AI is scheduled to publish the full open weights of Kimi K3 on Hugging Face at 00:00 UTC on July 27, which is this evening in the United States (8 p.m. Eastern, 6 p.m. Mountain). The model is a 2.8 trillion parameter mixture-of-experts system with a 1 million token context window and native vision, and the download runs to roughly 1.4 terabytes in MXFP4 4-bit quantization, a figure coverage has billed as the largest open-weight release to date. As of publication, the weights are not yet live.

The numbers behind the drop

Kimi K3 routes each token through 16 of its 896 experts, which works out to about 50 billion active parameters, according to an inference economics breakdown by Techi. At full 16-bit precision the weights swell to roughly 5.6 terabytes; the 1.4 terabyte figure reflects the MXFP4 quantization Moonshot is shipping. Until tonight, K3 has been available only through Moonshot's hosted API at kimi.com since its July 16 unveiling, with hosted access reportedly priced at $3 per million input tokens and $15 per million output tokens.

A hardware floor most teams cannot clear

Native MXFP4 execution requires NVIDIA's Blackwell generation or AMD's MI400 accelerators. Techi calculates that simply loading the model takes about eighteen 80 gigabyte accelerators before any memory is set aside for context or concurrent requests, and notes that a single node of eight 192 gigabyte cards can barely hold the weights. TechTimes has reported that Moonshot recommends 64 or more accelerators in a single interconnect domain for serving. Whatever the exact configuration, self-hosting K3 is a data center project, not a workstation download.

The license question

Coverage of the release, including TechTimes, has described the weights as arriving under a Modified MIT license. Techi cautions that the license terms have not actually been published and that, by Moonshot's own framing, the final text will land together with the weights themselves. Teams planning commercial deployment should read that document before pulling a single shard, because the difference between MIT and Modified MIT usually lives in the modifications.

Why self-hosting matters anyway

TechTimes frames the release around data sovereignty: an enterprise that self-hosts K3 keeps prompts, outputs and logs inside its own infrastructure, which no hosted API can match, and which matters particularly for organizations wary of routing sensitive workloads through a China-based provider. That argument only applies to the small population of companies that can actually field the hardware, but for them the calculus changes tonight, at least on paper.

A release under two spotlights

The drop comes at a loaded moment for Moonshot. The company filed for a Hong Kong listing earlier this month, as GaaS News covered in our report on the IPO filing, and a headline-grabbing open release is a familiar move ahead of a listing. It also arrives under government scrutiny: a joint assessment by the UK AI Security Institute and the US Center for AI Standards and Innovation, published ahead of the release and covered in our story on the K3 cyber assessment, found that the model's safeguards permitted offensive cyber attempts. Publishing the weights moves those behaviors from a hosted service, where Moonshot can intervene, to any data center with eighteen accelerators and a download script.

GaaS News will verify the actual artifacts, the final license text and the true download sizes once the Hugging Face repository goes live. Until then, every figure above describes a scheduled release, and schedules have slipped before.