Kimi K3's 2.8 Trillion-Parameter Weights Go Public July 27, Enabling Self-Hosting

Moonshot AI has set July 27, 2026 at 00:00 UTC for the public release of Kimi K3's full model weights, moving the 2.8-trillion-parameter model from API-only access to self-hostable open weights under a Modified MIT license.

• 2.8T total parameters, Mixture-of-Experts with 16 of 896 experts active per token (about 50B active), native 1M-token context
• Weight files run about 594GB in MXFP4 format; production self-hosting needs a supernode-class setup with 64+ accelerators and roughly 1.4TB of fast memory
• Moonshot's hosted API price already jumped to $3 per million input tokens (cache miss) and $15 per million output tokens, up from $0.95/$4 on predecessor K2.6
• Independent testing ranks K3 fourth on the Artificial Analysis Intelligence Index (57) while it leads Arena's preliminary WebDev coding leaderboard

Timing

The release lands a week after the White House accused Moonshot of covertly distilling Anthropic's Fable model and using banned Nvidia GB300 chips to train K3, claims outside researchers have challenged given that Fable only launched July 1 and K3 followed just over two weeks later. Self-hosting sidesteps a separate concern raised in that dispute: enterprises can now run K3 without routing data through Moonshot's China-based API. Compare it yourself in testingmodels' coding arena.