Kimi K3's Open Weights Are Out: Now Multimodal and Downloadable
We covered Kimi K3 when it was announced; back then there were no weights and Moonshot said the model was "behind the leaders." Now the open weights are actually out — with two new wrinkles: the model has gone multimodal, and its card carries some bold numbers.
In brief
- Weights are downloadable: Kimi K3 is now on Hugging Face under its own "Kimi K3 License." 2.8 trillion total / 104 billion active parameters (16 of 896 experts selected), 1 million token context. It's had ~99K downloads in a month.
- New: it handles images. A vision encoder called MoonViT-V2 (401M) was added — the announcement was mostly text-focused. It runs on engines like vLLM and SGLang, with weights in a compressed (MXFP4) format.
- Card scores (Moonshot's own measurements): GPQA Diamond 93.5; Terminal-Bench 2.1 88.3; BrowseComp 91.2; DeepSWE 67.5.
- Note: There's a tension between the company calling it "behind" at announcement and the strong numbers on the card — independent tests will settle it.
Our take
This is a 2.8-trillion-parameter model; running it on your own machine isn't practical for a regular user, so there's no "install it and go" win here. What matters is the direction: the open weights of a model claiming near-frontier performance are now out under an open license — and it handles images too. That has an indirect but real upside — open weights mean providers can host it cheaply and you aren't locked to a single vendor, which can lower your costs on coding or agent work down the line. Just treat the card numbers with caution: they're the company's own tests, not independent. For now the honest line is: open, ambitious, and now multimodal — but wait for independent comparisons before calling it "the best."
Kaynak: Moonshot AI — Hugging Face