- Sources: Hugging Face weights, repository, Moonshot post, HN 49065752, HN 49070985
- Summary: The Hugging Face models API read at this run returns moonshotai/Kimi-K3 as public and not gated, last modified 2026-07-27 at 16:29 UTC, with 5,192 likes, 2,850 downloads, 96 safetensors shards, about 1.56 TB of stored weights, and safetensors metadata totalling 2,779,931,837,184 parameters. The GitHub repository MoonshotAI/Kimi-K3 was created 2026-07-27 at 08:01 UTC and last pushed at 15:24 UTC, at 690 stars and 44 forks. Its README describes a Mixture-of-Experts model with 2.8T total and 104B activated parameters, 93 layers composed of 69 Kimi Delta Attention layers and 24 Gated MLA layers, 896 experts with 16 selected per token and 2 shared, a 160K vocabulary, a 1,048,576-token context, a MoonViT-V2 vision encoder of 401M parameters, and MXFP4 weights with MXFP8 activations under quantization-aware training. Licensing is a custom Kimi K3 License rather than an OSI license, carried on Hugging Face as license:other, so this is an open-weight release and not an open-source one. The benchmark table in the README is Moonshot's own and is not independently reproduced: it places Kimi K3 ahead of the named closed models on some rows and behind on others, and reports 88.3 on Terminal-Bench 2.1, 67.5 on DeepSWE, and 43.5 on HLE-Full. The release submission reached 985 points and a separate technical report submission reached 278 points at this run. The technical report itself was not read here, so nothing above rests on it.
- Why it matters: A downloadable 3T-class frontier model changes what a team can run on owned hardware, and it resolves the open-weight claim this digest has tracked since 2026-07-16, which earlier runs today could not confirm.
- Follow-up: Watch for independent benchmark reproduction, for the technical report's contents, and for how the custom Kimi K3 License constrains commercial deployment.
send feedback on this story