Kimi K3's 2.8T open weights drop at 00:00 UTC on July 27
The biggest open model yet leads Arena's code board — but serving it wants ~1.4TB of memory, its license is still unpublished, and it runs at one pricey tier.
Copy markdown
Frontier scale, now downloadable
Moonshot posts Kimi K3's weights to Hugging Face at 00:00 UTC July 27: a 2.8T-parameter sparse MoE (896 experts, 16 active per token, so ~50B effective compute) with a 1M-token context. It's the first frontier-tier open model at this scale — though Moonshot calls it a commitment, and open-weight dates have slipped before.
The catch: ~1.4TB of memory
The MXFP4 4-bit weights are ~594GB on disk but need ~1.4TB of fast memory to actually serve — roughly eighteen 80GB accelerators, or a single 8×Blackwell/MI400 node that barely fits them. 16-bit runs ~5.6TB. This is a datacenter drop, not a laptop one.
It leads on code — at one expensive gear
K3 tops Arena's Frontend/Code board (~1,679 Elo) and beats Claude Opus 4.8 max and GPT-5.5 high on knowledge work per Artificial Analysis, while trailing Claude Fable 5 and GPT-5.6 Sol. The snag: it exposes only a 'max' reasoning tier — Simon Willison's pelican SVG burned 13,241 tokens and 25¢.
License still unpublished
Moonshot hasn't posted K3's license, so commercial use stays unconfirmed until it ships alongside the weights. K2 used a Modified MIT that required attribution once you passed 100M users or $20M revenue — budget for a similar string before you build on it.
Can't self-host? Rent it today
If eighteen GPUs is out of reach, K3 already runs via the OpenAI-compatible API at platform.kimi.ai and free on the kimi.com web app. Self-hosting's main payoff is keeping inference traffic off Moonshot's servers — it doesn't cover training or the company's legal exposure.