Moonshot AI reports Kim K3.1 Pro tops rivals with 64.4 on finance test

What it claims

Moonshot AI released Kim K3.1 Pro on 23 October 2026 as an upgrade of Kimi K3, which it shipped in July 2026. All benchmark figures are Moonshot's own. The model leads Claude Opus 5.5 and GPT-6 Astra on one finance benchmark (OfficeQA Pro v2, 64.4). It sits between them on Terminal-Bench 4.0 and trails both on 3D modeling and cybersecurity. A separate Highspeed variant generates responses two to three times faster than the standard model.

How it works

The base is Kimi K3, a mixture-of-experts model with 896 expert networks, of which 16 are active at a time to cut inference cost. Kimi K3 uses Kimi Delta Attention, a custom implementation of Gated DeltaNet that reduces the RAM the attention mechanism needs to process prompts. K3.1 Pro adds post-training for complex reasoning and multi-step, tool-using tasks. Moonshot says this post-training used substantially more compute than its earlier projects. The training system also gained automated quality checks and infrastructure efficiency improvements.

Evidence

• Terminal-Bench 4.0 (agentic programming): 65.7%, against GPT-6 Astra at 58.2% and Opus 5.5 at 66.4%.
• Kimi Code Bench 3.0 Long horizon (internal): 61.2%, up from Kimi K3's 55.95%.
• BenchCAD (3D model coding): 92.7%, against Opus 5.5 at 96.2% and Astra at 95.9%.
• OfficeQA Pro v2 (financial document analysis, retrieval, numerical reasoning): 64.4, ahead of both rivals. Their scores are not given.
• Internal finance evaluations: financial modeling improved 7.2 points and financial research 11.6 points over Kimi K3.
• Kimi Hallu Bench Hard (internal): hallucination fell from 23.4% to 8.3%. In the same comparison, Opus 5.5 scored 15% with tools.

Moonshot also reports better generation of interactive environments, websites and 3D assets in Unity, Unreal Engine and Godot. One demonstration was a walking simulator with 296 skeletal and 402 muscle components. The model ran 167 development tasks and 420 reviews before the character walked continuously for 11.2 seconds.

How to use it

Both the standard and Highspeed editions are available now through paid access on Kimi.ai, Kimi Work, Kimi Code and the Moonshot API. Moonshot plans to open-source the full weights by 4 November 2026 for self-hosting. The page gives no licence, pricing, API model identifiers or hardware requirements.

Limits

Moonshot deliberately left certain cybersecurity data and tasks out of post-training to limit offensive capability. On an internal vulnerability-research evaluation, the released model scored 81%, against 90.5% for a version trained without those restrictions. Its cybersecurity benchmark results trail Astra and Opus 5.5.

Moonshot cautions that the Hallu Bench Hard test covers difficult scenarios and does not represent everyday hallucination rates. Several key results come from internal benchmarks that the page does not describe. The article names the model "Kim K3.1 Pro" but its predecessor "Kimi K3", so the exact product name should be checked against Moonshot's own release.