Moonshot publishes Kimi K3 weights: 2.8T parameters

Moonshot AI published the complete weights for Kimi K3 on 27 July 2026, eleven days after launching it as a hosted service. The Hugging Face repository holds 96 shards, the licence and the technical report.

K3 is a Mixture-of-Experts model with 2.8 trillion total parameters and 104 billion active. Sixteen of 896 routed experts fire per token alongside two shared experts. Context is 1,048,576 tokens. Quantisation-aware training runs from the supervised fine-tuning stage using MXFP4 weights and MXFP8 activations.

Moonshot recommends at least 64 accelerators. A four-bit estimate puts weights alone near 1.4TB. The licence is a custom Kimi K3 License; reports describing it as MIT are inaccurate.

Pricing is $3.00 per million cache-miss input tokens, $0.30 cache-hit, and $15.00 per million output, flat across the context window. Claude Fable 5.1 lists at $10 and $50.

The weekend after the release, Alibaba announced a 2.4-trillion-parameter Qwen 3.8 with open weights, a class it had previously kept API-only. Anthropic's February distillation allegation against Moonshot remains unresolved.

Moonshot acknowledges three constraints. The model is sensitive to thinking-history truncation, so harnesses that trim chain-of-thought degrade output quality. It shows excessive proactiveness in ambiguous scenarios. And conversational polish trails Claude Fable 5 and GPT-5.6 Sol.

K3 launched with a single reasoning effort level running at maximum thinking by default; low and high modes were promised in later updates. Most detailed scores are vendor-reported using mixed agent harnesses, so public weights improve inspectability without normalising the tests.

K3 ranked first on four of eight real-world task automation benchmarks, including Automation Bench, SpreadsheetBench 2 and BrowseComp, where it posted a state-of-the-art 91.2 out of 100.