Moonshot AI launches Kimi K3, a 2.8T parameter open MoE model with vision support

Last week, Beijing based AI startup Moonshot AI launched Kimi K3, a 2.8 trillion parameter Mixture of Experts model described as the first open model in its class and the largest open weights system announced so far. Kimi K3 uses 896 routing experts but activates only 16 per token to reduce computational costs. It also includes native vision support and a context window of up to 1 million tokens.

The model is available through Kimi’s web and desktop apps, Kimi Code, and Moonshot AI’s API. Pricing is $0.30 per million tokens for cached input, $3 for uncached input, and $15 for output. Moonshot plans to release the full model weights by July 27, 2026. Demand exceeded the company’s computing capacity within days, forcing it to temporarily suspend new subscriptions while expanding its infrastructure. Existing users were not affected.

Kimi K3 scored 81.2 on FrontierSWE, ahead of GPT 5.6 Sol and Claude Opus 4.8, and posted leading results across several software, automation, document, and visual reasoning benchmarks. Although Moonshot acknowledged that it still trails Claude Fable 5 and GPT 5.6 Sol in overall user experience, independent evaluations found it competitive with those proprietary frontier models. Its growing popularity have even caught the attention of the US government, with the Trump administration reportedly considering restrictions or a possible ban on Kimi K3 and other Chinese AI models over national security and intellectual property concerns, which could also apply to Alibaba’s recently previewed Qwen3.8 Max, another 2.4 trillion parameter Chinese model that the company claims ranks behind only Anthropic’s Fable 5.