Moonshot AI released the full weights of Kimi K3 for public download on July 27, roughly ten days after first unveiling the 2.8-trillion-parameter model through API access on July 16-17. The staggered rollout — benchmarks and hosted access first, downloadable weights second — was announced in advance by Moonshot and closes out the model's launch as the largest open-weight system released to date.

Scale built for sparsity, not for every GPU cluster

Kimi K3 is a sparse mixture-of-experts model that activates only 16 of its 896 experts per token, roughly 1.8% of its total parameters, alongside a 1-million-token context window, native multimodal (vision) input, and an always-on "thinking mode" reasoning path. The public weights are distributed in MXFP4 quantized format, cutting storage from what would be roughly 5.6 terabytes at FP16 down to about 1.4 terabytes — still enough to require multi-node clusters of 8-16 servers with 8 GPUs apiece (H100- or B200-class) for anyone looking to self-host rather than call Moonshot's API.

Where it lands against US frontier models

Moonshot's own published benchmarks put Kimi K3 ahead of Anthropic's Opus 4.8 and OpenAI's GPT-5.6 Sol on several evaluations, while it continues to trail Anthropic's higher-end Fable 5 model overall. Independent coverage has focused less on any single benchmark and more on the pace of the closing gap between Chinese open-weight releases and US closed frontier systems — commentary that has repeatedly invoked comparisons to the market reaction following DeepSeek's releases in 2025. Demand around the initial API launch was heavy enough that Moonshot temporarily paused new Kimi subscriptions in the days before the open-weight drop.

Why it matters

Full open weights change who can actually build on Kimi K3: enterprises and researchers can now fine-tune or self-host the model instead of depending on Moonshot's hosted API, trading a steep infrastructure bill for control over data and customization. The release also lands amid separate, unresolved friction between Washington and Moonshot over export-restricted Nvidia chips and disputed distillation allegations — controversies that have not slowed the model's rollout or its adoption by developers evaluating it as an alternative to US-built systems. Combined with recent releases from Z.AI and DeepSeek, Kimi K3's open-weight debut reinforces a pattern in 2026 of Chinese labs pairing near-frontier capability with permissive licensing, keeping pressure on how OpenAI, Anthropic, and Google price and gate access to their own top models.