Kimi K3 open weights drop: 2.8T params, self-host on 8x H100
The biggest open-weight model yet beats Opus 4.8 on several coding benchmarks, ships with vLLM day-0, and carries a license that bites above $20M revenue.
Copy markdown
What dropped: 1.56TB of weights on Hugging Face
Moonshot AI released Kimi K3's open weights on July 26 (~7:30pm EDT) — 2.8 trillion total parameters, 104B active per token, and a 1M-token context window, at 1.56TB on Hugging Face. The hosted API had been live since July 16; now the model itself is yours to pull.
It beats Opus 4.8 on coding, trails Fable 5 on general smarts
K3 posts 88.3% on Terminal-Bench 2.1 and tops the Frontend Code Arena, outscoring Claude Opus 4.8 on most coding tasks. On broad reasoning it trails Claude Fable 5 and GPT-5.6 Sol — so it's a coding/agent specialist, not an all-rounder.
Self-hosting is real, but heavy: 8x H100 minimum
vLLM supports K3 on day zero, but you need at least an 8x H100 80GB cluster — single-GPU inference is off the table. The payoff: run it inside your own VPC and keep prompts off a China-hosted API entirely.
The license catch: 'open weight,' not open source
K3 ships under a custom 'open weight' license, not MIT or Apache. Hobby and small-startup use is fine, but any entity above ~$20M annual revenue must sign a separate commercial agreement with Moonshot before shipping on it.
No cluster? The API is cheap
Don't want to stand up a cluster? Third-party providers serve K3 at roughly $3/M input (cache miss), $0.30 cache-hit, and $15/M output — a fraction of frontier coding-model pricing, which makes it a viable cheap default in a model-routing setup.