Kimi K3.1 Pro is Moonshot AI's most capable open-weight model to date. On October 23, 2026, Moonshot AI introduced Kimi K3.1 Pro on the official Kimi blog: the first model in the K3.1 series: a 3T-class model with a 1-million-token context window, built for coding, 3D generation, financial research and other agentic capabilities. Available now on Kimi.com, Kimi Work, Kimi Code and the Kimi API, with full weights to be released by November 4.
Coding results: 65.7% on Terminal-Bench, ahead of GPT-6 Astra
Coding leads the official scorecard: on Terminal-Bench 4.0 (strict tests of coding agents in terminal environments), K3.1 Pro scored 65.7%, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%); on the in-house Kimi Code Bench 3.0 long-horizon tasks it reached 61.2%, up from K3's 55.95%. An open-weight model now sits in the top tier of proprietary models on real terminal work.
3D and finance: two underrated tracks
3D generation is the other headline: 92.7% with tools on BenchCAD (rebuilding 3D from multi-view renderings via CAD code), close to Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%); the concurrently launched 3D Make lets users preview and edit 3D models on kimi.com and the Kimi app and export them for printing. In finance, 64.4 on OfficeQA Pro v2 (multi-step numerical reasoning over real financial documents), claimed state of the art; the hallucination rate on Kimi Hallu Bench Hard fell from 23.4% (K3) to 8.3%.
How it was built: K3 with scaled-up post-training
K3.1 Pro is essentially a post-trained K3: substantially larger post-training compute, plus large-scale joint reinforcement learning across hundreds of downstream domains. Moonshot AI says it optimized the RL framework, long-context inference, training efficiency and sandbox infrastructure so each increment of RL compute delivers stable, predictable gains. A Highspeed variant ships alongside, 2-3x faster for more responsive interaction.
The open-weight tradeoff: safety subtraction before release
Notably, Moonshot AI curbed capability before releasing weights, removing cyber-related data from mid-training and RL. On its internal xdev-bench-hard test (vulnerability discovery and exploitation), the final K3.1 Pro scored 81.0%, below the checkpoint without data removal (90.5%) and below frontier proprietary models public for at least three months. The stated principle: open weights should not push the frontier of offensive cyber capability further, while keeping general usefulness for defensive work.
Three things worth watching next
First, the weight release before November 4: the license, quantized versions and community fine-tuning ecosystem decide whether it can replicate a DeepSeek-style open-weight moment. Second, Highspeed's real pricing and API performance — whether 2-3x speed comes with quality trade-offs. Third, terminal coding agents are becoming the must-win track for open weights; with the bar raised, who follows next?
