The Key Facts
Kimi K3 first went live via API and apps on July 16: a Mixture-of-Experts model with 2.8 trillion total parameters, roughly 16 of 896 experts active per token, a 1-million-token context window, and native vision input, priced at about $3 per million input tokens and $15 per million output. At launch, the company also pre-announced that the full weights and technical report would land in Hugging Face's `moonshotai` org at 00:00 UTC on July 27—the milestone at the center of this story.
Two Things to Watch
First, the license. K3 shipped without a model card or a license file, so exactly how "open" these open weights are only becomes clear once the weights and card publish; the K2 line used a modified MIT license, but that shouldn't be assumed for K3. Second, benchmark validation. K3 topped the Arena.ai Frontend Code Arena as an anonymous checkpoint (1679, ahead of Claude Fable 5 and GPT-5.6 Sol), but scores like Terminal-Bench 2.1 have so far been Moonshot's own numbers, pending the technical report and independent retesting shipping alongside the weights.
The Bar, and Why It Matters
If it lands on schedule, this is the first time the open-source camp puts a 3-trillion-parameter-class model on the table—both a new option and a new barrier for teams running their own inference: community estimates put even quantized versions at roughly 650GB–1TB of memory and multi-node deployment (some reports say MXFP4 quantization can compress it to about 1.4TB). It leads on frontend coding, precisely where recent proprietary flagships have each shown gaps—another sign the open/closed gap is narrowing month by month on specific capabilities.
To Be Confirmed
As of publication, no Kimi-K3 repository has appeared in the official `moonshotai` Hugging Face org, whose newest entry is still June's Kimi-K2.7-Code. The actual go-live time, file formats, and license are whatever the official org page shows; any "K3 weights download" surfacing before the official repo is likely a third-party placeholder or misinformation and should not be trusted.
via: Moonshot AI official launch announcement (2026-07-16, pre-announcing the July 27 weights and technical-report release); Hugging Face `moonshotai` org page; verified 2026-07-27