kimi k3 weights are out — 2.8 trillion parameters, modified MIT, and a 51% hallucination rate you should know about

Moonshot AI published the Kimi K3 model weights on Hugging Face on July 27, 2026 under a Modified MIT license. at 2.8 trillion parameters, K3 is the largest open-weight model ever released — by a significant margin. the previous largest was around 600 billion parameters. this is a different tier.

the architecture is Kimi Delta Attention (KDA), a hybrid linear attention mechanism with Attention Residuals. native visual understanding and a 1M-token context window ship with the base weights. the technical report drops alongside the weights.

the benchmark picture (self-reported)

Moonshot AI's own evals put K3 ahead of Claude Opus 4.8 max and GPT-5.5 high on most coding benchmarks, while trailing Claude Fable 5 and GPT-5.6 Sol. on the Frontend Code Arena — where K3 scored 1,679 points at announcement — it ranked first, ahead of Fable 5 at 1,631. on SWE-Bench Verified it posts 68.4%, compared to Fable 5's 80%+ range.

pricing for the hosted API is $3 per million input tokens and $15 per million output tokens — positioned above Grok 4.5 ($2/$6) but below Claude Opus 4.8 ($15/$75). for teams self-hosting on the open weights, inference cost is whatever your compute costs.

the number moonshot didn't highlight: 51% hallucination rate

Artificial Analysis's independent AA-Omniscience benchmark measured K3's hallucination rate at approximately 51%. the previous generation, Kimi K2.6, was at 39%. both numbers are higher than most frontier models — Fable 5 and GPT-5.6 Sol score well below 30% on the same benchmark.

moonshot did not include hallucination metrics in their published evaluation. the AA-Omniscience finding comes from third-party testing and was not disclosed or contested by Moonshot as of the weights release date. for coding use cases where the model is generating code that will be run in a controlled environment, hallucination rate matters less than on tasks where factual accuracy in prose is critical. for agentic workflows that mix coding with research, documentation lookup, or decision-making, a 51% hallucination rate is a real risk to account for.

what the open weights actually unlock

the Modified MIT license allows commercial use without restrictions, model fine-tuning, and redistribution. the 2.8T parameter size is a practical constraint: running K3 at full precision requires significant multi-GPU infrastructure. quantized versions (Q4/Q8) will reduce memory requirements but come with quality trade-offs that haven't been independently benchmarked yet.

the more immediate unlock is for researchers and teams who want to inspect or fine-tune at scale. an open-weight 3T-class model creates new possibilities for domain-specific fine-tuning on coding tasks — something that wasn't possible at this capability tier before today.

honest take

K3 is a genuine milestone on open weights. the Frontend Code Arena performance is real, the context window is real, and the license is genuinely permissive. the hallucination numbers are also real, and the fact that they weren't in the launch materials is worth noting. if you're evaluating K3 for agentic use, benchmark it on your specific tasks rather than assuming the self-reported coding numbers generalize to your pipeline.