Moonshot AI 發佈 Kimi K3.1 Pro，K3.1 系列首款模型

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

According to the announcement, K3.1 Pro Highspeed delivers two to three times the speed of K3.1 Pro with identical model capabilities. 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 Macs; in Kimi Code, where users select the model with the /model command; and through the Kimi API, where the models are listed as kimi-k3.1-pro and kimi-k3.1-pro-highspeed, with Zero Data Retention available for API customers. Kimi Enterprise administrators can enable K3.1 Pro for their organizations, with access off by default at launch.

Training and Reported Performance

K3.1 Pro is a post-trained version of Kimi K3, built with what Moonshot AI describes as substantially larger post-training compute. To enable large-scale joint reinforcement learning across hundreds of downstream domains, the company said it optimized every layer of the stack: the reinforcement learning framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, 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 at 58.2 percent and narrowly behind Claude Opus 5.5 at 66.4 percent. On Kimi Code Bench 3.0 Long-horizon, an internal benchmark covering multi-step execution, tool use, and end-to-end problem solving, K3.1 Pro scored 61.2 percent against K3’s 55.95 percent. On BenchCAD, which tests 3D reconstruction from multi-view renderings through CAD code generation, K3.1 Pro scored 92.7 percent with tools, approaching Claude Opus 5.5 at 96.2 percent and GPT-6 Astra at 95.9 percent. A 3D Make feature is live on kimi.ai and the Kimi app, letting users preview and edit 3D models online and download them ready for 3D printing.

Moonshot AI also reports a state-of-the-art score of 64.4 on OfficeQA Pro v2, which tests complex fiscal data analysis over real financial documents, and cites gains over K3 across its internal finance benchmarks: 7.2 points on Financial Modeling, 11.6 points on Financial Research, and about 4.2 times K3’s performance on Business Operation. On Kimi Hallu Bench Hard, an internal stress test for factual accuracy, the model’s claim-level hallucination rate fell from 23.4 percent for K3 to 8.3 percent, while the share of responses containing at least one factual error was 42.4 percent of its predecessor’s rate. The blog notes that the evaluation is deliberately crafted to probe error-prone situations and that the reported rates do not represent overall error rates in daily use.

In one internal long-horizon coding test described in the blog, K3.1 Pro built a human walking dynamics simulator on top of human-atlas, an open-source anatomical model of the human body comprising 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews, the model iteratively refined a walking controller; when reward shaping failed to improve arm motion, it ran open-loop probes of the simulator’s dynamics, identified and corrected 36 errors in shoulder-muscle moment-arm calculations, and retrained the controller.

Cybersecurity Controls and Open-Weight Principles

Moonshot AI said it observed vulnerability-research performance improving steadily as post-training scaled, and traced the trend to cyber-related content embedded within the general mid-training corpus and cyber-related tasks in the reinforcement learning environment mix. For the open-weight release of K3.1 Pro, the company said it identified and removed cyber-related data from both training stages while retaining the general-purpose coding tasks needed for model usefulness.

On xdev-bench-hard, the company’s in-house benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks, a checkpoint trained without those risk-reduction measures solves 90.5 percent of problems at pass@3 while the final K3.1 Pro reaches 81.0 percent, the blog reports. Moonshot AI said the results show data removal meaningfully reduces cyber capability relative to what the same training pipeline would otherwise produce, and that K3.1 Pro’s cyber capability remains behind that of frontier proprietary 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 cyber capability. Second, it intends to progressively raise the cyber capability available in open weights, stating 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 Configuration in Kimi Code

Separate Kimi Code documentation details how the models are exposed to developers. The docs list six model IDs across the K3.1 Pro, K3, K2.8 Preview, and K2.7 Code Highspeed families, describe k3.1-pro as a 3T-parameter flagship with a context window of up to 1,048,576 tokens for higher-tier members, and describe the Highspeed version as carrying the same capabilities with about twice the quota usage. K3.1 Pro is fully rolled out in Kimi Code, according to the page.

According to the docs, K3.1 Pro models support thinking mode only, and requests with thinking disabled are rejected with an HTTP 400 error. Both new models accept reasoning_effort settings of low, high, or max, with max as the default, and both accept image and video input. Under new membership 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; under legacy plans, the corresponding tiers are Moderato and Allegretto.

The documentation recommends starting a new session rather than switching to the new models mid-session, using the native Responses, chat completion, or message APIs for the most stable output, and entering exact model IDs in third-party tools, because a mistyped ID falls back to K2.8 Preview without thinking.

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