Kimi K3, live on PrivateMind

A few days ago we wrote about Kimi K3, Moonshot AI's 2.8-trillion-parameter open-weight model, and said Options Technology was evaluating it. K3 is live on PrivateMind, behind the same gateway, with the same audit trail, on the same infrastructure as every other model we serve.

In the first few hours after launch, users at Options Technology processed 3 billion tokens through it.

What it takes to run

A 2.8-trillion-parameter model does not run on a spare server. Moonshot recommends supernode configurations with 64 or more accelerators; with the shipped MXFP4-quantised weights, the floor for a single serving instance is one HGX B300 node, before any of the optimisation work begins. Getting from 'the weights load' to 'the weights serve well' took KV-cache tuning to keep long contexts economical across the 1-million-token context window, and performance work across prefill and decode. Compatibility engineering with both SGLang and vLLM means K3 drops into the same serving stack as everything else on the platform.

That gap, between an open-weight model existing and an open-weight model running well inside a controlled estate, is most of what PrivateMind is for: the weights may be open, but serving them properly is the work.

Try it

If you have a PrivateMind account, K3 is available today. Log in at privatemind.com and pick it from the model list. If you are not set up yet, use the sign-up form at the top of this page, or on the website, and someone will be in touch.

PrivateMind is an enterprise product, and SSO is the standard way organisations sign in. The form is where that conversation starts.

The same question

Every post on this blog comes back to the same question: where the weights sit, and who can reach them. For K3 the answer is the same as for every other model on the platform. The model runs on our GPUs, on our network, with no external dependencies anywhere in the data path, and the only journey a prompt makes is from your team to the gateway and back again.