Kimi K3.1 Pro

Moonshot AI introduced Kimi K3.1 Pro as a post-trained successor to its open-weight Kimi K3 model in an October 24, 2026 Kimi blog announcement. The launch post emphasizes benchmark gains, but the available analysis cautions that it is a launch announcement rather than a technical report and therefore provides limited methodological detail. A distinctive part of the announcement was Moonshot’s frank discussion of reducing the model’s hacking capabilities. That safety decision makes the release relevant to developers evaluating a coding-focused model whose cyber-related abilities were deliberately trimmed before public release.

Moonshot AI released Kimi K3.1 Pro as a heavily post-trained successor to Kimi K3, retaining a 1-million-token context window while increasing reinforcement-learning compute across hundreds of downstream domains. Moonshot highlighted coding, 3D creation, and financial research as major improvement areas, with agentic-coding results reportedly within one point of Claude Opus 5.5. The model launched in the Kimi apps, Kimi Code, and the Kimi API, with full weights scheduled for November 4, 2026. DataCamp also reports that Moonshot deliberately removed cyber-related training data before the open-weight release, a notable safety choice for a model aimed at long-horizon coding and agentic work.

Eigent's technical overview describes Kimi K3.1 Pro as Moonshot AI's newest open-weight model, a post-trained version of Kimi K3 announced on October 23, 2026, aimed at full-project coding, parametric 3D modeling, and end-to-end financial workflows. It keeps K3's sparse mixture-of-experts design with 2.8 trillion total and 104 billion active parameters, plus multimodal input and a 1-million-token context window. The piece attributes the gains to large-scale joint reinforcement learning across hundreds of downstream domains, with the RL framework, long-context inference, and sandbox infrastructure tuned so each increment of RL compute delivers stable returns, and notes the Highspeed variant runs 2-3x faster. Reported benchmarks include Terminal-Bench 4.0 at 65.7% versus GPT-6 Astra at 58.2% and Claude Opus 5.5 at 66.4%, Kimi Code Bench 3.0 Long-horizon at 61.2% versus K3's 55.95%, BenchCAD at 92.7% with tools, and OfficeQA Pro v2 at 64.4.

NVIDIA's Build model card documents Kimi K3.1 Pro as a 3T-class multimodal model from Moonshot AI for reasoning, coding, and agentic workflows, accepting text, images, and video and generating text within a 1,048,576-token context window. The card explicitly states the model is not owned or developed by NVIDIA and remains governed by the Kimi K3.1 Pro License. Deployment on Build.NVIDIA.com is dated 11/04/2026, with global deployment geography. The architecture is listed as a transformer mixture-of-experts with 2.8 × 10^12 total parameters, 1.04 × 10^11 active parameters, a 160,000 vocabulary size, and Kimi K3 as the base model. Documented use cases include code generation, image and video understanding, retrieval-augmented generation, document question answering, tool-assisted reasoning, CAD scripting, and financial research, with Moonshot recommending images up to 4096 × 2160 pixels and video up to 1920 × 1080 pixels.

Moonshot AI's tech blog introduces Kimi K3.1 Pro as the first model in the K3.1 series and its most capable 3T-class open-weight model, with a 1-million-token context window and substantial improvements in coding, 3D creation, and financial research. It is a post-trained version of Kimi K3 using substantially larger post-training compute, with the RL framework, long-context inference, training efficiency, sandbox infrastructure, and rollout scheduling all optimized. The same post introduces Kimi K3.1 Pro Highspeed, a distinct variant delivering 2-3x the speed of K3.1 Pro for more responsive interactions, and states both variants are available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API with full weights slated for release by Nov 4. Moonshot reports 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra at 58.2% and just behind Claude Opus 5.5 at 66.4%, plus 61.2% on its internal Kimi Code Bench 3.0 Long-horizon versus K3's 55.95%.

Moonshot AI's official model card describes Kimi K3.1 Pro as an open-weight, natively multimodal agentic model and the first release in the K3.1 series, post-trained from Kimi K3 through large-scale joint reinforcement learning across hundreds of downstream domains. It pairs that with a 1,048,576-token context window and gains in coding, 3D creation, financial research, and factual accuracy. The card lists a mixture-of-experts design with 2.8T total parameters and 104B activated, 93 layers, 896 experts selecting 16 per token, and attention split between 69 KDA and 24 Gated DSA layers, where Gated DSA replaces the full-attention Gated MLA layers of Kimi K3 for long-context efficiency. Weights are MXFP4 with MXFP8 activations, and text plus image are supported via a 401M-parameter MoonViT-V2 vision encoder. Reported results use reasoning effort max, temperature 1.0, and top-p 0.95, with Terminal-Bench 4.0 at 65.7.