Kimi K3 open weights land: 2.8T MoE, 1M context, day-0 hosting
Moonshot's 104B-active model is live on Fireworks, OpenRouter and vLLM — strong at agentic and coding work, but a custom license and quirks need reading first.
Copy markdown
What you can run today
Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 104B active per token, a 1M-token context window, and text, image and video input. The weights are on Hugging Face — pull them for Transformers, vLLM or SGLang, or call it hosted.
The cost pitch vs Opus 5
On OpenRouter it runs $3 in / $15 out per million tokens across six providers. Fireworks benchmarks it at 92.7% on SWE tasks for $0.52 each against Opus 5's 94.8% at $1.05 — it claims up to 5x better cost-per-task.
Where it's actually strong
Fireworks ranks it #1 in the world on front-end code and #3 on both DeepSWE and the Artificial Analysis Intelligence Index (57). It posts 91.2 on BrowseComp, edging Fable 5 and GPT-5.6 Sol — sharpest at agentic and browser work.
'Open weights' with an asterisk
It doesn't ship under Apache or MIT. It's a custom 'Kimi K3 License' that mirrors proprietary-style terms for research and commercial use. Read it before you build a product on top — downloadable is not the same as unrestricted.
Plan for the quirks
Moonshot flags the model as highly sensitive to its thinking history and 'excessively proactive,' and warns against switching models mid-session. Budget for explicit behavioral constraints before it goes near production.
Build this weekend
104B active params means a serious GPU box, not a laptop — but with 1M context and open weights you can stand up an owned agentic coding or research assistant with no per-token meter. Day-0 vLLM recipes are already published.