Kimi K3.1 Pro: What's New, Benchmarks, and How It Compares

Kimi K3.1 Pro is Moonshot AI's newest open-weight model — a post-trained version of Kimi K3 that the company calls its most capable 3T-class open-weight model, with a 1-million-token context window. Announced on October 23, 2026, it targets real agentic work: full-project coding, parametric 3D modeling, and end-to-end financial workflows. This guide covers what K3.1 Pro is, its headline benchmarks, how it compares to K3 and the leading proprietary models, and when the open weights land.

What Is Kimi K3.1 Pro?

Kimi K3.1 Pro is the first model in Moonshot AI's new K3.1 series and a post-trained version of Kimi K3, built with substantially more post-training compute. It keeps K3's sparse Mixture-of-Experts (MoE) design — 2.8 trillion total parameters with 104 billion active per token — along with multimodal input and a 1-million-token context window, according to the Kimi K3.1 Pro tech blog.

Moonshot says the gains come from large-scale joint reinforcement learning across hundreds of downstream domains, with the RL framework, long-context inference, and sandbox infrastructure tuned so each increment of RL compute delivers stable returns. The result, per Moonshot, is consistent improvement in every domain over K3.

A faster variant, K3.1 Pro Highspeed, runs 2–3× faster than the standard model for more responsive, interactive use.

Kimi K3.1 Pro Benchmarks

Moonshot positions K3.1 Pro as matching or approaching the strongest proprietary models across coding, agentic, and knowledge benchmarks. The headline numbers from the tech blog:

• Terminal-Bench 4.0 (coding agents on realistic, resource-constrained terminal tasks): 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%).
• Kimi Code Bench 3.0 Long-horizon (internal multi-step execution and tool use): 61.2%, up from K3's 55.95%.
• BenchCAD (3D reconstruction from multi-view renderings via CAD code): 92.7% with tools, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
• OfficeQA Pro v2 (complex fiscal data analysis over real financial documents): 64.4, which Moonshot reports as state of the art.

Treat internal and launch-day benchmarks as a starting point rather than a settled verdict — scores are reported by the vendor and evaluation harnesses differ between models.

Coding: Full-Project Builds and Game Engines

K3.1 Pro covers the full range of coding work, from quick fixes in real codebases to complete, from-scratch projects. Moonshot highlights that it works directly inside game engines — Unity, Unreal Engine, and Godot — and combines them with web search and 3D generation to build complete experiences, with stronger composition, typography, and animation in front-end pages and Three.js scenes.

One internal test shows the long-horizon ambition: Moonshot reports that K3.1 Pro built a human walking-dynamics simulator on top of an open-source anatomical model, and across 167 missions and 420 reviews, iteratively refined a controller until the simulated human walked continuously for 11.2 seconds — at one point running open-loop probes of the simulator's dynamics and correcting 36 errors in shoulder-muscle calculations rather than blindly tweaking rewards.

3D Modeling and Parametric Design

Spatial reasoning is a standout. Moonshot says K3.1 Pro excels at parametric modeling — mechanical parts, architectural models, configurable products, and custom furniture — and can write and run scripts for tools like Blender and CAD software without MCP connectors to build and refine assets, edit components, and adjust geometry and materials.

From a text brief, sketch, or reference image, the model infers shape, proportions, and spatial relationships to produce usable 3D models. Moonshot also launched 3D Make on kimi.ai and the Kimi app, which lets you preview and edit models online and export them for 3D printing or as video.

Finance: Fewer Hallucinations, More Reliable Modeling

Finance is where Moonshot leans hardest on reliability. On internal benchmarks, K3.1 Pro improves over K3 by 7.2 points on financial modeling and 11.6 points on financial research, and reaches roughly 4.2× K3's performance on a simulated business-operation benchmark.

The accuracy story is the one that traveled furthest. On Kimi Hallu Bench Hard — a stress test deliberately crafted to probe errors — K3.1 Pro's claim-level hallucination rate fell from 23.4% for K3 to 8.3%. Moonshot is explicit that this is a stress test and does not represent overall error rates in everyday use, so read the "65% fewer hallucinations" framing as a benchmark delta, not a guarantee.

Cybersecurity: A Deliberately Capped Open-Weight Release

Because the weights will be public, Moonshot describes targeted controls on offensive-security capability. It says it removed cyber-related data at both the mid-training and RL stages before the open-weight release while keeping general coding tasks intact.

The effect is measurable on its internal xdev-bench-hard: a checkpoint without those controls solved 90.5% of tasks at pass@3, while the final K3.1 Pro reaches 81.0%. Moonshot states that K3.1 Pro's offensive cyber capability remains behind frontier proprietary models that have been public for at least three months, and frames gradual, controlled diffusion as better for defensive security than concentrating the capability in closed systems.

Kimi K3.1 Pro vs. K3

If you've used K3, K3.1 Pro is a post-training upgrade on the same architecture rather than a new base model — same 2.8T MoE and 1M context, with broad agentic gains layered on top.

The practical difference: K3.1 Pro is tuned for long-horizon coding, 3D/CAD, and finance work, with notably lower hallucination on stress tests and a faster Highspeed option for interactive use.

How It Compares to Claude and GPT-6

On the benchmarks Moonshot published, K3.1 Pro lands close to the frontier without clearly passing it. It sits just behind Claude Opus 5.5 on Terminal-Bench 4.0 (65.7% vs 66.4%) and ahead of GPT-6 Astra (58.2%); on BenchCAD with tools it trails both (92.7% vs 96.2% and 95.9%). The headline is that an open-weight model is now trading blows with proprietary leaders on agentic coding and CAD — the same trajectory we tracked with Meta's Muse Spark Open and Reflection Beam.

Availability and Open Weights

Kimi K3.1 Pro and K3.1 Pro Highspeed are available now across the Kimi product family — on kimi.ai and the Kimi mobile apps, Kimi Work (desktop app v3.1.0+), Kimi Code (select the model with /model), and the Kimi API (kimi-k3.1-pro / kimi-k3.1-pro-highspeed). Enterprise access is off by default at launch.

The key date for self-hosters: full model weights release by November 4, 2026. Until then, K3.1 Pro is API- and app-only; after that, teams can inspect, modify, and run it on their own infrastructure.

Do This With Your Own AI Workforce

A model like K3.1 Pro earns its keep when it's wired into real workflows — repos, 3D pipelines, financial documents, and multi-step plans — not just a chat box. A model-agnostic, multi-agent platform lets you route long-horizon coding to a frontier open-weight model while keeping sensitive data local. Eigent is an open-source Cowork desktop app that runs a multi-agent AI workforce on your own machine, so you can plug in the best model per task. See how agents can review GitHub PRs end to end, then download Eigent to try it.

Frequently Asked Questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is Moonshot AI's open-weight model announced October 23, 2026 — a post-trained version of Kimi K3 with 2.8 trillion total parameters (104B active), a 1-million-token context window, and multimodal input. Moonshot calls it its most capable 3T-class open-weight model, with improvements in coding, 3D creation, and finance.

Is Kimi K3.1 Pro open source?

Moonshot plans to release the full model weights by November 4, 2026, at which point developers can inspect, modify, and self-host it. Until then, K3.1 Pro is available only through Kimi apps, Kimi Work, Kimi Code, and the Kimi API.

How does Kimi K3.1 Pro compare to Claude Opus 5.5 and GPT-6 Astra?

On Moonshot's published benchmarks, K3.1 Pro scores 65.7% on Terminal-Bench 4.0 — ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%). On BenchCAD with tools it scores 92.7%, trailing both Opus 5.5 (96.2%) and Astra (95.9%). These are vendor-reported numbers with differing harnesses.

What's new in K3.1 Pro versus K3?

K3.1 Pro is a post-trained K3 on the same 2.8T MoE architecture and 1M context, with broad agentic gains: stronger long-horizon coding, better 3D and CAD work, and finance improvements of 7.2 points (modeling) and 11.6 points (research). Its stress-test hallucination rate drops from 23.4% to 8.3%, and a Highspeed variant runs 2–3× faster.

Can I use Kimi K3.1 Pro with a multi-agent platform like Eigent?

Yes. Once the weights ship, a model-agnostic platform like Eigent can route long-horizon coding and knowledge-work tasks to K3.1 Pro while using other models for routine steps — and keep data local when needed.