Moonshot AI launches Kimi K3.1 Pro, the first model of the new K3.1 series – Unite.AI

Moonshot AI introduced Kimi K3.1 Pro on October 24, 2026, the first model in its new K3.1 series, along with a faster variant called Kimi K3.1 Pro Highspeed. The company describes the K3.1 Pro as its most capable 3T-class open-weight model to date, with a 1 million token contextual window.

According to the announcement, K3.1 Pro Highspeed offers two to three times the speed of K3.1 Pro with the same features as the model. Both models are available across the Kimi product family: on Kimi.ai and the Kimi mobile apps for iOS, Android and HarmonyOS; on the Kimi Work desktop application, version 3.1.0 or later, for Windows and Apple Silicon Mac; in Kimi Code, where users select the model with the /model command; and through the Kimi API, where models are listed as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with Zero Data Retention available to API customers. Kimi Enterprise administrators can enable K3.1 Pro for their organizations, with access disabled by default at launch.

Training and performance reported

K3.1 Pro is a post-workout version of Kimi K3, built with what Moonshot AI describes as substantially larger post-workout compute. To enable large-scale joint reinforcement learning across hundreds of downstream domains, the company said it has optimized every layer of the stack: the reinforcement learning framework, long-context inference, training efficiency, sandbox infrastructure, launch planning, and a shared platform for environment quality checks.

The company reports that on Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks, K3.1 Pro scored 65.7 percent, ahead of GPT-6 Astra with 58.2 percent and just behind Claude Opus 5.5 with 66.4 percent. On Kimi Code Bench 3.0 Long-horizon, an internal benchmark that covers multi-phase execution, tool usage, and end-to-end troubleshooting, K3.1 Pro scored 61.2% versus K3’s 55.95%. On BenchCAD, which tests 3D reconstruction from multi-view rendering through CAD code generation, the K3.1 Pro scored 92.7% with the tools, coming close to Claude Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. A 3D Make feature is live on kimi.ai and the Kimi app, allowing users to preview and edit 3D models online and download them ready for 3D printing.

Moonshot AI also reports a cutting-edge score of 64.4 on OfficeQA Pro v2, which tests complex tax data analyzes on real financial documents, and cites gains over K3 in its internal financial benchmarks: 7.2 points on financial modeling, 11.6 points on financial research, and approximately 4.2 times K3’s performance on business operations. In the Kimi Hallu Bench Hard, an internal stress test for factual accuracy, the model’s statement-level hallucination rate dropped from 23.4% for K3 to 8.3%, while the percentage of answers containing at least one factual error was 42.4% of its predecessor’s rate. The blog notes that the rating is deliberately crafted to probe error-prone situations, and that the reported rates do not represent overall error rates in everyday use.

In a long-horizon internal coding test described in the blog, K3.1 Pro built a human walking dynamics simulator on top of the Human Atlas, an open-source anatomical model of the human body comprising 296 skeletal components and 402 muscular components. Through 167 missions and 420 independent reviews, the model iteratively refined a walking controller; when reward modeling failed to improve arm movement, he performed open-loop probes of simulator dynamics, identified and corrected 36 errors in shoulder and arm muscle moment calculations, and retrained the controller.

Cyber ​​security controls and open weight principles

Moonshot AI said it observed a steady improvement in performance in vulnerability research as it scaled after training, and that it trended toward cyber-related content integrated into the overall mid-training corpus and cyber-related tasks into the mix of the reinforcement learning environment. For the open-weight version of K3.1 Pro, the company said it identified and removed cyber-related data from both training phases, while retaining the generic coding tasks necessary for the model’s usefulness.

On xdev-bench-hard, the company’s internal benchmark of expert-curated, source-based vulnerability detection and exploitation efforts, a checkpoint trained without such risk mitigation measures resolves 90.5% of issues at pass@3 while the final K3.1 Pro achieves 81.0%, the blog reports. Moonshot AI said the results show that data removal significantly reduces computing capacity compared to what the same training pipeline would otherwise produce, and that K3.1 Pro’s computing capacity remains lower than that of proprietary frontier models that have been publicly available for at least three months.

The blog frames two principles behind the open weight release. First, the company aims not to release open weights that would extend the absolute frontier of computing capability. Second, it intends to progressively increase the computing capacity available in open weights, saying that non-discriminatory access gives more teams the means to audit the code they maintain, including legacy systems that rarely receive dedicated security reviews.

Model setup in Kimi code

The separate Kimi Code documentation describes in detail how models are exposed to developers. The documents list six model IDs in the K3.1 Pro, K3, K2.8 Preview, and K2.7 Code Highspeed families, describe k3.1-pro as a 3T-parameterized flagship with a context window of up to 1,048,576 tokens for higher-level members, and describe the Highspeed version as having the same features with approximately double the quota usage. K3.1 Pro is fully implemented in Kimi Code, according to the page.

According to the documents, the K3.1 Pro models only support thinking mode, and requests with thinking disabled are rejected with an HTTP 400 error. Both new models accept low, high, or maximum reasoning_effort settings, with maximum as the default, and both accept image and video input. With the new subscription plans, k3.1-pro requires Plus and above, with the 1M context window reserved for Pro and above, while k3.1-pro-highspeed requires Pro and above; in legacy plans, the corresponding levels are Moderato and Allegretto.

The documentation recommends starting a new session rather than switching to new templates mid-session, using native replies, chat completion, or message APIs for more stable output, and entering exact template IDs into third-party tools, because a mistyped ID falls into the K2.8 preview without thinking about it.

Full model weights for the Kimi K3.1 Pro will be released by November 4, 2026, the company said.