Moonshot launches Kimi K3.1 Pro: 2.8T MoE open-weight flagship (1M ctx); API today, full weights by Nov 4; AA Index 55 (#4)

Why it matters

【Key points】

Moonshot (@Kimi_Moonshot) launched Kimi K3.1 Pro at ~17:15 UTC on 2026-10-23: most capable 3T-class open-weight model to date per vendor — MoE 2.8T total / 104B active, 1M context, multimodal. Live today on Kimi.com, Work, Code, and API (plus Highspeed, claimed 2–3×). Full weights by Nov 4, 2026. Blog: https://www.kimi.com/blog/kimi-k3-1-pro

Vendor / third-party figures (attribute): Terminal-Bench 4.0 65.7% (vendor: above GPT-6 Astra 58.2%, just under Claude Opus 5.5 66.4%); BenchCAD with tools 92.7%; OfficeQA Pro v2 64.4. Artificial Analysis same day: Intelligence Index 55 (+11 vs K3), 4th overall, leading open-weights once published; GDPval-AA Elo 1846 (3rd); ~33% cheaper than K3 input/output. Vendor also documents cyber-data cleaning (xdev-bench-hard pass@3 90.5% → 81.0%).

【Verifiable】

https://www.kimi.com/blog/kimi-k3-1-pro

https://x.com/Kimi_Moonshot/status/2113680701770068067

https://x.com/Kimi_Moonshot/status/2113681863109939271

https://x.com/ArtificialAnlys/status/2113688083145314416

【Our read】

1. Same size as K3 with heavier post-train / joint RL — not a new parameter class. Product now, weights ~Nov 4: API-first then open, opposite recent DeepSeek/Qwen “weights first” cadence.

2. Split vendor Terminal-Bench from AA Index. AA’s #4 / pending open-weight lead is meaningful third-party signal; many coding benches still use vendor-chosen harnesses — don’t read as a clean Opus wipeout.

3. Coding / 3D-CAD / finance-agent teams get a high-tier open-weight path to trial via API; price cut vs K3 with similar $/Index per AA. Cyber section is more discussable than score-only blogs, still vendor-run.

4. Limits: weights not out; internal benches unverifiable; Highspeed vs Pro gap underspecified; “open-weight” claim is forward-looking until Nov 4. Don’t treat AA #4 as beating all closed flagships.