The launch and access methods
月之暗面 (Moonshot AI) announced Kimi K3.1 Pro, the first model in the K3.1 series and, according to the company, the most powerful it has ever built: a 3-trillion-parameter model with a context window of up to 1 million tokens. The news, reported by IT之家, indicates that the release occurred overnight between October 23 and 24, 2026.
The model is available immediately via kimi.com, the Kimi app, Kimi Work, Kimi Code, and the API on the developer platform. Alongside the standard version, Kimi K3.1 Pro Highspeed makes its debut, presented as equivalent in terms of capabilities but with a declared output speed two to three times higher; however, the cost for API use is double. The model's full weights will be published by November 4, an element that signals the intention to make the model inspectable and reusable, although the material does not specify the license.
The benchmarks: claimed comparisons and internal tests
According to the official description cited by the source, on coding, agent, and knowledge benchmarks K3.1 Pro reaches or approaches the levels of the strongest closed models. On Terminal-Bench 4.0, which evaluates programming agents in terminal tasks with limited resources and task-specific tests, the model scores 65.7%, above the 58.2% attributed to GPT-6 Astra and close to Claude Opus 5.5's 66.4%. On Kimi Code Bench 3.0 Long-horizon, an internal benchmark covering multiple engineering domains and testing multi-step execution, tool use, and end-to-end solution, the score is 61.2% versus 55.95% for K3.
The most interesting declared capability concerns 3D creation and spatial reasoning. On BenchCAD, which requires generating CAD code to reconstruct 3D models from multi-view renders, K3.1 Pro scores 92.7% with tool use, versus 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra. The company claims that the model can derive shape, proportions, and spatial relationships from text, a sketch, or a reference image, and that it can directly write and execute scripts for Blender and CAD tools without MCP connectors, useful in parametric modeling of mechanical parts, buildings, and furniture. A 3D Make feature has also been introduced on kimi.com and in the app, with online preview and editing, download for 3D printing, and export to video or live photo.
Finance, reliability, and what remains to be verified
On the financial front, the source reports a leaderboard-topping result: 64.4 on OfficeQA Pro v2, an external benchmark based on real financial documents that requires information retrieval, cross-document verification, and multi-step numerical reasoning. On the three internal benchmarks, the company reports improvements over K3: +14% in Financial Modelling, +20% in Financial Research, and a score about 4.2 times that of its predecessor in Business Operation, which simulates management and decision-making over long horizons.
On reliability, in the internal stress test Kimi Hallu Bench Hard, the error rate on individual claims falls from 23.4% for K3 to 8.3%, while the share of responses containing at least one factual error is stated to be about 42.4% of that of the previous model. The same source warns that this is a test built around cases where models err most often and that the values do not reflect everyday use.
Several points remain to be clarified. Most of the comparisons are based on data provided by the company, and several benchmarks are internal and therefore not independently replicable; the results on Terminal-Bench and BenchCAD, although referring to external tests, are not corroborated by third parties in the material. The exact publication date of the weights and the license are not indicated, and the hardware requirements for local execution are not specified. The real scope of the financial improvements will need to be assessed in operational scenarios outside the controlled tests.
