Kimi K3.1 Pro nears frontier with 65.7% Terminal-Bench score, as Moonshot deliberately trims its exploit-writing skills
maier_ak · x · 2026-10-26
Moonshot AI released Kimi K3.1 Pro, a post-trained successor to the 2.8T-parameter Kimi K3 MoE, with a 1M-token context window and substantially more RL compute across hundreds of task domains; a Highspeed variant claims 2-3x faster inference. It's live in apps and API, but weights are only promised by Nov 4 — until then no outside verification is possible.
Key claims
- Terminal-Bench 4.0: 65.7%, vs GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4%
- BenchCAD (3D reconstruction as CAD code): 92.7%
- Several gains are on internal benchmarks
The unusual part: deliberate capability trimming
The launch post candidly describes measuring and reducing the model's ability to write working exploits before open-weight release: exploit skill grows via post-training because security content sits in general data, and removing it cut an internal exploit benchmark from 90.5% to 81.0% (pass@3). The author values the rare with/without comparison but notes 81% remains high for open weights, and the claim that it trails closed models can't yet be checked externally.
More from Models
- Lyra neuron explorer: steering Qwen 3.5 4B activations turns it into a 'money-making' model — rudrank · 2026-10-26
- Harness switch reveals underrated Gemini 4 Argon, now second only to Opus 5.5 — banteg · 2026-10-26
- Leak: Kimi K3.1 Flash reportedly in training, DeepSeek 4.1 Pro still 'grinding' — teortaxesTex · 2026-10-26
- Editing just 2 neurons instantly "flips" Qwen 3.5 4B, netizens joke — jaivinwylde · 2026-10-26
- ChatGPT Pro user with 95% quota left hit with "capabilities reduced" notice — Shozab_haxor · 2026-10-26
- Google auto-converts one-off searches into AI Mode conversations, users frustrated — Status_Technician364 · 2026-10-26