Moonshot AI's Kimi K3 debuts as China's largest open-weight model

Moonshot AI has released Kimi K3, a 2.8 trillion-parameter language model that now stands as the largest open-weight system to come out of China. The model was announced on July 16, 2026, with weights promised by July 27 and subsequently published to Hugging Face on schedule.

At 2.8 trillion total parameters, Kimi K3 overtakes DeepSeek's 1.6T v4 Pro as the heaviest openly released model from a Chinese lab. Some early coverage rounded the figure to "3T-class," but Moonshot and independent trackers both confirm the precise count sits at 2.8 trillion. The full weight files run to 1.56TB.

Moonshot is pointedly calling this an "open weight" release rather than "open source," a distinction that matters given the licensing terms attached. Unlike Kimi K2's modified MIT license from July 2025, K3 ships under a proprietary license that requires attribution from any entity with over 100 million monthly active users or more than $20 million in monthly revenue, and mandates a separate commercial agreement for Model-as-a-Service operators crossing $20 million in aggregate revenue over 12 months.

Performance and pricing shift

Benchmark results suggest the model closes ground on Western frontier systems. In LMArena's Frontend Code Arena, K3 posted a 76% pairwise win rate, ahead of Claude's Fable 5 at 63% and GPT-5.6 Sol at 58%, landing the model at #1 with 1,679 points, a dramatic jump from K2.6's #18 ranking.

Moonshot attributes part of that gain to architectural efficiency, claiming roughly 2.5 times more intelligence delivered per unit of compute, alongside a 21% drop in output tokens needed relative to K2.6.

Kimi K3's cyber capabilities lag far behind those of frontier models, according to the UK's Cyber Institute.

That efficiency claim doesn't extend everywhere. The-Decoder reports that K3's math performance trails frontier models in independent testing, and the UK's Cyber Institute found its cyber capabilities notably weaker than leading Western systems.

Pricing has jumped sharply alongside capability. K3 costs $3 per million input tokens and $15 per million output tokens, a steep rise from K2.6's $0.95/$4 rates. OpenRouter now lists the model through seven separate providers at matching prices.

Beyond the model itself, Moonshot open-sourced supporting infrastructure, including high-performance attention kernels, a mixture-of-experts communication library, and tooling for running AI agents at scale. That infrastructure push, paired with a genuinely competitive coding benchmark showing, signals Moonshot is trying to compete on more than raw parameter count as the frontier race between US and Chinese labs continues to narrow.