Kimi K3
The world's first open 3T-class model — with a license that isn't quite either
What it is
Moonshot AI's 2.8T-total / 104B-active MoE (896 experts, top-16, 93 layers mixing KDA and gated MLA) with a 1M-token context, shipped natively in MXFP4 via quantization-aware training — a 1.56TB download. Multimodal in, agentic by design, with day-0 hosting across seven OpenRouter providers.
Why it's interesting
The most-downloaded debut of the summer and the top-scoring open-weights model on the Artificial Analysis index — 88.3 Terminal-Bench 2.1, 93.5 GPQA-Diamond, 91.2 BrowseComp. But 'open weight' is Moonshot's phrase, not ours: the bespoke Kimi K3 License hangs real conditions on it, and this site exists to read those before you build.
Use cases
- Long-horizon coding and research agents with 1M-token context
- Self-hosted frontier-class inference for teams under the license thresholds
- Fine-tuning and quantization research at trillion scale
Who it's for
Infra-rich labs and inference providers — after a license read
Setup
Advanced. ~1.4TB resident memory at native MXFP4; realistically a multi-node cluster or a hosted provider
Limitations & cautions
The hardware bill is the largest of any open-weights release ever, the license needs a lawyer once you're successful, and a ~51% hallucination rate on AA-Omniscience says trust-but-verify still applies at 2.8T parameters.
Editorial takeaway
Last week we wrote 'verify the actual LICENSE file before celebrating.' We did. Celebrate the engineering; read clause by clause.