Skip to content

SuperQode v0.2.55: Kimi K3 on every host serving the open weights

Choose a tag to compare

View all tags
@Shashikant86 Shashikant86 released this 27 Jul 17:56
v0.2.55
fcabe91

Moonshot released the Kimi K3 open weights on 26 July 2026: 2.8T parameters, 1M context, MXFP4. Because the weights are open, the same model is now served by several providers — each under a different id. SuperQode ships the verified id for every one of them, so you pick a provider and the model just works.

Supported routes

Provider Model id
moonshot kimi-k3
baseten moonshot-ai/Kimi-K3
fireworks accounts/fireworks/models/kimi-k3
together moonshotai/Kimi-K3
openrouter moonshotai/kimi-k3
siliconflow moonshotai/Kimi-K3
vllm, sglang moonshotai/Kimi-K3-MXFP4

Every id was checked against that host's own model page. Note that Baseten spells the publisher moonshot-ai with a hyphen while the others use moonshotai — copying an id between providers will not work.

Getting started

export BASETEN_API_KEY=...
superqode

:connect byok baseten moonshot-ai/Kimi-K3

Or run :connect and pick Baseten under Model Hosts. Any provider above works the same way with its own key. The same behaviour on every route Whichever host serves it, SuperQode applies the kimi model pack — maximum reasoning and parallel tool calls so switching providers changes latency and price, not how the agent behaves.

New providers

  • Baseten : first-class OpenAI-compatible host (BASETEN_API_KEY), a day-0 host for K3 with 1M context and native vision. Set
    BASETEN_BASE_URL to point at a dedicated deployment.
  • Modal : serverless GPUs. Modal gives you your own deployed endpoint rather than a shared API, so set MODAL_BASE_URL to the URL your deployment exposes.

Self-hosting

vllm and sglang both have day-0 support, with a KDA prefix-caching implementation contributed upstream by the Kimi team. This is a datacenter exercise rather than a laptop one: the MXFP4 weights are roughly 594GB and Moonshot recommends 64 or more accelerators.

Assets 2
Loading