Kimi K3.1 Pro open weights land Nov 4 — 2.8T params, 1M context
Moonshot trades blows with Opus 5.5 on agentic coding and ships with cyber skills capped — and the drop lands as China moves to license 2T+ weight exports.
Copy markdown
Trades blows with Opus 5.5 — on Moonshot's own scorecard
On Moonshot's self-reported benchmarks, K3.1 Pro hits 65.7 on Terminal-Bench 4.0 — 0.7 behind Claude Opus 5.5's 66.4 — and actually beats it on SWE Marathon v1.1, 51.7 to 46.3. But it trails badly on Program Bench, 31.5 to 65.0. These are vendor numbers, so wait for independent leaderboards before betting a pipeline on it.
Use it now via the API, self-host from Nov 4
It's live as kimi-k3.1-pro on an OpenAI-compatible endpoint at roughly $2/$10 per million input/output tokens with a 1M-token context, so you can point Claude Code or any OpenAI client at it today. Full open weights are promised by Nov 4 — but the base is K3's 2.8 trillion parameters, so self-hosting is a cluster job, not a laptop one.
The weights ship with cyber skills dulled on purpose
Moonshot stripped cyber data from mid-training and the RL environments: on its internal xdev-bench-hard exploit suite the released model scores 81%, down from 90.5% for an unmitigated checkpoint (K3 scored 33.3%). It says it won't open-weight frontier cyber capability — a rare case of an open-weight lab self-capping offensive skills before shipping.
China is building a license gate for exactly this kind of drop
MOFCOM's Announcement No. 48 (Oct 16) would require an export license for frontier weights above 2T parameters or 10^26 training FLOP — and Kimi K3 is 2.8T. Already-public weights and plain API/SaaS use are exempt, so your self-hosted Qwen and DeepSeek stay fine, but future flagship open-weight releases could get gated. Public comment closes Nov 15.