Update: vLLM ships production serving for Kimi K3's open weights
The 2.8T open model dropped at 00:00 UTC — now here's how to run it: a 16-GPU serving config, the 64+ GPU reality, and the rent-the-API math.
Copy markdown
vLLM ships day-one production serving
vLLM landed production-scale K3 support to launch with the weights: fused KDA and MLA kernels across the 93-layer network, MXFP4 MoE, validated on a 16-GPU DP16+EP16 configuration, with both NVIDIA and AMD paths. That's what turns a 1.4TB download into a live endpoint.
The hardware wall: still 64+ accelerators
Moonshot's own deployment guidance calls for 64+ accelerators to serve the full 2.8T model — datacenter-scale, not a workstation. The MXFP4 weights alone run ~1.4TB, so vLLM's 16-GPU validation is a floor, not the full-throughput target.
Can't self-host? Rent it at $3/$15 per M
platform.kimi.ai lists K3 at $3.00 input and $15.00 output per million tokens, with a $0.30 cache-hit rate — the realistic path for most builders until third-party providers list the model.
Read the LICENSE, not the tweets
License terms weren't finalized in public as the drop landed; earlier Kimi models shipped under a Modified-MIT. If you're building commercially, check the actual LICENSE file on the repo before you ship on it.