Kimi K3.1 Pro and K3.1 Pro Highspeed

977 points · 584 comments on HN · read original →

Points and comments are a snapshot, not live.

Kimi K3.1 Pro matches or approaches frontier proprietary models on coding, 3D, and finance benchmarks.

Kimi K3.1 Pro is a 3T-class open-weight model with a 1-million-token context window, available today. A Highspeed variant offers 2-3x faster response. On Terminal-Bench 4.0 it scores 65.7%, ahead of GPT-6 Astra (58.2%) and behind Claude Opus 5.5 (66.4%). On BenchCAD for 3D reconstruction it scores 92.7% with tools. On OfficeQA Pro v2 for finance it scores 64.4. The model's cyber capability is deliberately reduced by removing cyber-related training data and RL environments, resulting in 81.0% pass@3 on an internal exploit benchmark versus 90.5% without those controls. Full weights release by Nov 4, 2026.

The company also details improvements in factual accuracy, with the hallucination rate on a stress-test benchmark dropping from 23.4% to 8.3%.

What commenters are saying

Commenters are impressed by the benchmark results, noting the model appears frontier-level. A key point of discussion is Moonshot's deliberate reduction of cyber capabilities in the open-weight release. Many support this restraint, arguing the world needs time to prepare and that open Pandora's box would cause havoc. Others argue that defense requires offense, and that nerfing the model only benefits governments while leaving defenders without adequate tools. Several commenters note the model still lags behind frontier proprietary models on cyber tasks, and that GLM-5.3 remains the go-to for cybersecurity. One commenter highlights the significant improvement in hallucination rate from K3 to K3.1 Pro.