The weights are out. As of today, July 27, 2026, Moonshot AI has released the full open checkpoint for Kimi K3 — the first open-weight model in the three-trillion-parameter class. If you have been waiting to download it and run it in your own VPC, you can. The question that matters now is not can you self-host it. It is should you, and the answer hides inside two numbers that the headline collapses into one. (If you're still deciding whether K3 belongs in your stack at all, start with the founder's guide to what K3 actually is; this piece is for the teams already asking how to serve it.)

The headline is 2.8T. The number that sets your bill is 50B#

Kimi K3 is a 2.8T-A50B model: 2.8 trillion total parameters, but a mixture-of-experts design that fires only about 50 billion of them per token. The router picks 16 of 896 experts for each token, so while the whole model sits resident in memory, the compute per token is that of a roughly 50-billion-parameter model.

That distinction is the whole story of serving K3, and it cuts two ways:

The 2.8T sets what you must store. The 50B active sets what you must compute. You pay for the first in GPUs and the second in tokens — and people conflate them constantly.

Store versus compute. Get that separation right and the rest of the decision falls out of it.

What "self-host" actually costs in hardware#

The open checkpoint is roughly 1.4TB in the MXFP4 format Moonshot trained and shipped. That 1.4TB is not your deployment size — it is the floor just to load the weights. You still need headroom for KV-cache and activations, and this is where K3's 1M-token context window turns into a hardware bill: under real concurrency, the KV-cache, not the weights, is what blows up.

A practical serving deployment lands near an 8-node cluster of 8×80GB GPUs — about 5TB of aggregate GPU memory. That is 64 top-tier accelerators to serve one model. This is a data-center deployment, not a workstation one, and no amount of "it's only 50B active" changes the fact that all 2.8T parameters must be resident because any expert can be selected next.

The serving shape#

You do not run a 2.8T MoE with a one-line docker run. You run it on a multi-node inference engine — vLLM or SGLang are the mainstream choices — pointed at the MXFP4 checkpoint, with a parallelism plan that shards the model to fit:

# Illustrative shape, not a guaranteed flag set — your engine's
# day-one K3 support decides the exact invocation.
vllm serve moonshotai/Kimi-K3 \
  --quantization mxfp4 \
  --tensor-parallel-size 8 \      # within each node
  --pipeline-parallel-size 8 \    # across the 8 nodes
  --enable-expert-parallel \      # shard the 896 experts
  --max-model-len 1000000

The download is the easy part. The real work is the distributed configuration: getting tensor-parallelism inside a node, pipeline- or expert-parallelism across nodes, and a KV-cache policy that survives your actual concurrency without OOM-ing on the million-token context. Budget your week for that, not for wget.

The one line that decides rent vs. own#

Here is the math nobody puts in the launch post. A 64-GPU cluster runs on the order of $90,000+ per month rented at commodity GPU-cloud rates — and that meter runs whether the GPUs are saturated or idle. The hosted Kimi K3 API costs $3 per million input tokens and $15 per million output tokens, and you pay only for tokens you actually send.

So the break-even is a utilization line:

Owning the cluster beats renting the API only when you keep ~64 GPUs near-saturated. Below that line, "open weights" lowered your license cost to zero and left your compute bill exactly where it was — often higher, because you're now paying for idle silicon.

For the overwhelming majority of founders and small teams, sustained saturation of a 64-GPU cluster is nowhere close. That does not make the open release worthless — it makes it a hedge and a residency option, not a default cost cut. Self-host K3 when you have (a) genuinely high steady volume, (b) data that legally cannot leave your boundary, or (c) a real need to never depend on Moonshot's API terms again. Otherwise, prototype on the hosted API — you can call Kimi K3 in about ten minutes — measure quality on your workload, and let real token volume, not the thrill of a 1.4TB download, tell you when to buy the cluster.

The weights being open is a milestone worth marking. Just don't let the free download talk you into a six-figure monthly GPU bill you didn't need.