Back to LLM Models

Moonshot's K3.1 Pro: near-frontier coding in a 3T open-weight model

Moonshot AI shipped Kimi K3.1 Pro on October 23, the first model in its K3.1 series and, by the company's own account, its most capable 3T-class open-weight model to date. It carries a 1-million-token context window. A companion variant, Kimi K3.1 Pro Highspeed, runs 2-3x faster according to Moonshot, aimed at more responsive interactions. Both are live today on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, and the full model weights are scheduled for release by November 4.

A post-training scale-up, not a new base model

K3.1 Pro is a post-trained version of Kimi K3, not a new base model. Moonshot says it applied substantially larger post-training compute and ran large-scale joint reinforcement learning across hundreds of downstream domains. The company describes optimizing the whole stack, from the RL framework and long-context inference to training efficiency, sandbox infrastructure, and rollout scheduling, so that each increment of RL compute returns stable, predictable gains. The stated result is consistent improvements in every domain it tested.

The benchmark picture

All the numbers below are Moonshot's, and that framing matters when you read them. On Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks with strict per-task tests, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and 0.7 points behind Claude Opus 5.5 at 66.4%. On Kimi Code Bench 3.0 Long-horizon, an internal benchmark for multi-step execution and tool use, K3.1 Pro reaches 61.2% against 55.95% for K3. The internal-benchmark caveat applies to that second pair: Moonshot sets the tasks and the harness.

Two data points land outside coding. On BenchCAD, which tests 3D reconstruction into CAD code, K3.1 Pro scores 92.7% with tools, approaching Claude Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). On OfficeQA Pro v2, complex fiscal analysis over real financial documents, Moonshot reports a state-of-the-art 64.4.

Where the capability shows up

The coding examples in the post range from quick fixes in real codebases to full-project builds. The model works across Unity, Unreal Engine, and Godot, and combines them with web search and 3D generation to create complete experiences. The most concrete internal test: K3.1 Pro built a human walking dynamics simulator on top of human-atlas, an open-source anatomical model with 296 skeletal components (including cartilage) and 402 muscle components. Across 167 missions and 420 independent reviews, it analyzed results, diagnosed problems, and iterated the walking controller until the simulated human went from collapsing within seconds to walking continuously for 11.2 seconds. When reward shaping stopped improving arm motion, the model ran open-loop probes of the simulator's dynamics, found and corrected 36 errors in shoulder-muscle moment-arm calculations, and retrained.

For finance, Moonshot reports gains across three internal benchmarks: financial modeling up 7.2 points over K3, financial research up 11.6 points, and a simulated business-operation task at roughly 4.2x K3's performance. On factual accuracy, the company's Kimi Hallu Bench Hard shows the claim-level hallucination rate falling from 23.4% for K3 to 8.3% for K3.1 Pro. Moonshot's own footnote: that benchmark is a stress test with prompts deliberately crafted to probe error-prone situations, so the rates are not daily-use error rates.

Kimi also launched 3D Make on kimi.ai, where generated 3D models can be previewed and edited online, downloaded ready for 3D printing, or exported as videos or Live Photos.

Cyber capability, deliberately capped

The policy-relevant part of the post is how Moonshot handled offensive cyber capability. The company found that vulnerability-research performance rose steadily as post-training scaled, and traced the source: cyber-related content in the general mid-training corpus and cyber tasks in the RL environment mix. For the open-weight release, it identified and removed cyber-related data from both training stages while keeping the general coding tasks it needs for usefulness.

On xdev-bench-hard, Moonshot's in-house benchmark of source-code-based vulnerability discovery and exploitation calibrated to its current capability level, the checkpoint without these measures solved 90.5% of tasks at pass@3. The shipped K3.1 Pro reaches 81.0%. Moonshot says the model's offensive cyber capability is below what its own pipeline could produce and trails frontier proprietary models that have been publicly available for at least three months. The company frames this as an open-weights policy: it does not want to extend the absolute cyber frontier in open weights, and it intends to raise that capability gradually, on the reasoning that teams with defensive needs can then audit code on infrastructure they control.

What to take from this

If you're choosing models for coding or agentic work, K3.1 Pro adds a new row to the decision table: near-frontier terminal-benchmark scores in an open-weight package, at a scale (3T parameters, 1M context) that earlier open-weight releases had not reached. You can start on the API today. Moonshot offers zero data retention for API customers, and enterprise administrators can enable the model for their organization, off by default at launch. When the weights land by November 4, self-hosting becomes the real test. The benchmark numbers are Moonshot's, so the move is to run your own workloads before you reprice anything.

Source: kimi.ai

More news

Back to LLM Models

Have something in mind?

Start a conversation