Kimi K3 Weights Are Now Public — One Day Early, and Despite a Sanctions Cloud
Moonshot AI released the full open weights llm files for Kimi K3 on July 26 at around 7:30 PM EDT — one day ahead of the July 27 date the company had communicated publicly, and in the middle of an unresolved US government accusation that the model was built by covertly distilling Anthropic's Fable 5.
The weights tip the scales at roughly 1.4 to 1.56 TB. Running them locally is effectively out of reach for almost everyone: the model's 2.8-trillion-parameter sparse MoE architecture — 16 of 896 experts active per forward pass — requires an estimated 2 TB of VRAM, which puts it squarely in multi-rack datacenter territory. Moonshot didn't offer a quantized or smaller-context variant at launch.
For most developers in the kimi ai community, the practical story is the hosted access that materialized within hours of the drop. Together AI and Modal both stood up inference endpoints on day zero, making K3 accessible via API without owning the hardware. Pricing for hosted inference from third-party providers varies, but K3's API price through Moonshot's own platform has been $3 per million input tokens and $15 per million output — competitive with US frontier mid-tier models but several times more expensive than K2.6, the prior Moonshot release.
The timing matters beyond the logistics. As we covered earlier this week, the White House accused Moonshot of building K3 through industrial-scale, covert distillation of Anthropic's models, and Treasury Secretary Scott Bessent put sanctions on the table as a possible response. Moonshot hasn't responded publicly to either accusation. Publishing the weights publicly doesn't resolve those questions — distillation accusations concern training data, not distribution format — but it does mean that whatever K3 contains is now available to any researcher who wants to inspect it, which cuts against the narrative of a model that needs to hide what it is.
