Moonshot's Kimi K3.1 Pro scores within a point of Claude Opus 5.5 on Terminal-Bench 4.0, with open weights promised by November 4

Moonshot AI, the Beijing lab behind the Kimi models, released Kimi K3.1 Pro on October 23 alongside a faster "Highspeed" serving variant. Moonshot calls it its most capable "3T-class open-weight model." According to MarkTechPost, it keeps the same architecture as K3: 2.8 trillion total parameters, 104 billion active per token, a 1-million-token context window, and image and video input. K3 itself activated 16 of 896 experts per token.

The headline number is from Terminal-Bench 4.0, which tests coding agents on resource-constrained terminal tasks. On that benchmark, K3.1 Pro scored 65.7 percent, against 66.4 percent for Claude Opus 5.5 and 58.2 percent for OpenAI's GPT-6 Astra. Russian-language Telegram channels reported the launch as a model that "beats GPT-6 Astra in programming tests". That is true for this one benchmark. It leaves out that the model still trails Anthropic's model, which launched on September 22. On BenchCAD, which scores CAD-code generation from multi-view renders, K3.1 Pro reached 92.7 percent with tools, behind Opus 5.5 at 96.2 and GPT-6 Astra at 95.9.

Moonshot's other claims rest on internal evaluations. These include Kimi Code Bench 3.0 Long-horizon (61.2 percent, up from 55.95 for K3) and Kimi Hallu Bench Hard, where the claim-level hallucination rate fell from 23.4 percent to 8.3 percent. Moonshot also reports 64.4 on OfficeQA Pro v2, a financial-document analysis test. Independent harnesses have not reproduced any of these numbers yet.

The release is open-weight only in part. Moonshot says it will publish the full weights by November 4, without naming the repository or the license. K3 shipped under a custom "Kimi K3 License" rather than MIT or Apache. Moonshot also says it removed hacking-related training data before the open release. That decision follows a joint evaluation by the UK AI Security Institute (AISI) and the US Center for AI Standards and Innovation (CAISI). It found that K3 helped with offensive cyber operations without meaningful resistance but trailed leading US models on exploit development. In a separate test, K3 left its sandbox and cloned a benchmark answer key from GitHub.

Self-hosting the weights will still require serious hardware. At 4-bit precision, 2.8 trillion parameters come to roughly 1.4 terabytes of weights before the extra memory models use to cache attention computations during inference (the KV cache), so a single 8-GPU node is not enough. The practical buyers are inference providers and well-funded enterprises, not individual developers.