- Sources: arXiv 2607.24653, technical report PDF, Hugging Face weights, HN 49070985
- Summary: Moonshot published the Kimi K3 technical report on arXiv as 2607.24653, with the same document carried in the MoonshotAI/Kimi-K3 repository. This is the follow-up the 2026-07-27 entry asked for. The report's benchmark table is Moonshot's own and is not independently reproduced: it places K3 ahead on SWE-Marathon at 42.0 and MCPMark-Verified at 94.5. The report's abstract concedes that K3 trails both Claude Fable 5 and GPT-5.6 Sol overall, with DeepSWE, FrontierSWE, and HLE-Full among the benchmarks where it is behind. The weights carry a custom Kimi K3 licence rather than the modified MIT licence used for K2, and ship as 96 safetensors shards at about 1.56 TB. The architecture the report documents was published with the weights on 2026-07-27: 2.8T total parameters with 104B active, 93 layers split into 69 Kimi Delta Attention layers and 24 Gated MLA layers, 896 experts with 16 selected per token plus 2 shared, a 401M-parameter MoonViT-V2 vision encoder, a 1,048,576-token context, and MXFP4 weights with MXFP8 activations under quantization-aware training.
- Why it matters: The report is the citable architecture document for a 2.8T-parameter open-weights model, and the licence change away from K2's modified MIT terms is what a team checks before committing to a deployment.
- Follow-up: Watch for independent benchmark reproduction, and for how the custom Kimi K3 licence constrains commercial deployment.
send feedback on this story