Kimi K3 open weights: 2.8T MoE, MXFP4, 1M context, video-native
Moonshot's 104B-active MoE now downloads in Safetensors — MXFP4 shrinks it to ~1.4TB, with day-0 hosting on Together and Modal; a custom license is the catch.
Copy markdown
2.8T total, 104B active per token
K3 is a sparse mixture-of-experts — 896 experts, 16 fired per token, 93 layers — so it runs at ~104B active while carrying 2.8T of capacity. It's natively multimodal, with a 401M-param MoonViT-V2 encoder covering images and video.
MXFP4 weights: ~1.4TB, not 5.6TB
Moonshot quantization-aware-trained the weights straight to MXFP4 (MXFP8 activations), cutting memory bandwidth ~4x vs FP16 and shrinking the download to roughly 1.4TB. That's what makes serving a 2.8T model merely expensive instead of impossible.
Strong on coding and reasoning
The official card lists GPQA Diamond 93.5 and DeepSWE 67.5; community evals put it 1st on Next.js and Frontend Code Arena. Treat the leaderboard as a starting point and re-run your own suite before you ship.
1M context, but check the license
The 1,048,576-token window swallows whole-repo digests and long multimodal prompts in one shot. Weights ship under a custom 'Kimi K3 License' — not the Modified-MIT of past drops — so read the LICENSE file before any commercial use.
Day-0 hosting; self-host wants a supernode
Together AI and Modal confirmed same-day hosted inference, and Moonshot serves it on Mooncake with a 90% cache-hit rate on coding. Local runs stay impractical — Moonshot recommends 64+ accelerators — so most builders will rent, not own.
The API, if you skip the weights
The kimi-k3 endpoint runs $0.30 cached / $3 input / $15 output per Mtok, with max-effort thinking on by default — so even short calls burn reasoning tokens. Cache your repo prefixes to keep landing that $0.30 rate.