Moonshot Releases Kimi K3 Model Weights on Hugging Face

Moonshot AI didn't ease into this one. Kimi K3 lands as a 2.8-trillion-parameter mixture-of-experts model, the biggest openly-released model class the world has seen, and Moonshot is calling it the first open 3T-class system built for frontier-level work. Only 104 billion of those parameters activate per token, thanks to a new Stable LatentMoE setup that fires up 16 of 896 experts at a time, which the company says delivers roughly 2.5x better scaling efficiency than its own Kimi K2.

The architecture underneath is genuinely new, not just bigger. Moonshot swapped in something called Kimi Delta Attention paired with Attention Residuals, mixing 69 KDA layers with 24 gated MLA layers across 93 total layers. There's also a MoonViT-V2 vision encoder bolted on natively, so the model reads text, images and video without a separate pipeline, and it can hold a million tokens of context — enough to chew through an entire codebase or a stack of research PDFs without losing the thread.

Moonshot is pitching K3 less as a chatbot and more as an agent you can leave alone for a while. The benchmarks back that framing up: it posts 88.3 on Terminal-Bench 2.1, beats every rival listed on SWE-Marathon at 42.0, and edges out Claude Opus 4.8 and GPT-5.5 on tool-heavy tasks like AutomationBench and MCP-Atlas. It doesn't win everywhere — Claude Fable 5 still leads on raw reasoning benchmarks like HLE-Full and CritPt — but K3 trades blows with models that, unlike K3, nobody can actually download.

There's a practical angle too. Moonshot trained K3 with quantization-aware training from the fine-tuning stage onward, running MXFP4 weights with MXFP8 activations, which should make a 2.8T model somewhat less absurd to actually serve on real hardware. It's already wired up to vLLM, SGLang and Moonshot's own TokenSpeed engine, plus an OpenAI/Anthropic-compatible API if you'd rather not self-host a multi-terabyte model in your garage.

The license is Moonshot's own Kimi K3 License rather than something like Apache or MIT, so read the fine print before you build a company on top of it. But the headline stands: a Chinese lab just shipped weights for a model that competes with the best closed systems from OpenAI and Anthropic on agentic coding and long-horizon tasks, and put them on Hugging Face for anyone to pull down.