Everyone's screenshotting the K3 benchmark table. The two sections worth reading are further down.
Pricing ｜ $0.30/MTok on a cache hit, $3.00 on a miss, $15.00 output.
Same input, 10x apart. On a 1M-token context model that stops being a billing footnote and becomes the main cost variable. Moonshot says their API holds a >90% cache hit rate on coding workloads — that number is in the launch post because without it, a 2.8T model has no unit economics to defend.
Limitations ｜ Ｗritten by the vendor.
Preserved-thinking mode needs the full thinking history passed back every turn; drop it, or switch mid-session from another model, and quality goes unstable. It's also proactive to a fault — give it ambiguous intent and it decides for you, so explicit constraints in the system prompt or AGENTS.md aren't optional. And they admit a UX gap versus Claude Fable 5 and GPT 5.6 Sol.
Publishing that third one on launch day takes some nerve. Together the three say something specific: open weights solved access, not serving.
Look at what running K3 actually asks of you. 2.8T params with MXFP4 QAT. Two caches — KDA recurrent state and MLA KV — managed jointly. Prefix caching that survives a million tokens. Thinking state held across turns and tool calls. Supernodes of 64+ accelerators. Miss one link in that chain and you're not running the model in the post.
Frontier capability is becoming a cost curve you can negotiate with. Capturing it is an engineering problem, not a download problem.