Kimi K3.1 Pro

DataCamp reports that Moonshot AI released Kimi K3.1 Pro, a heavily post-trained upgrade of Kimi K3 with the same 1M-token context, alongside a faster K3.1 Pro Highspeed variant, available in the Kimi apps, Kimi Code, and the Kimi API, with full weights following by November 4, 2026. Moonshot rebuilt its RL framework, sandbox infrastructure, and rollout scheduling so added compute produced steady gains across every measured domain. The analysis frames agentic coding as the headline result, with K3.1 Pro trading blows with Claude Opus 5.5 and beating GPT-6 Astra on terminal work, while pure reasoning and knowledge benchmarks still trail closed frontier models by a clear margin. It also notes that Moonshot stripped cyber-related training data before the open-weight release and that the API price is lower than Kimi K3's.

Eigen's technical overview says Moonshot AI announced Kimi K3.1 Pro on October 23, 2026 as the first model in its K3.1 series, a heavily post-trained version of Kimi K3 that keeps the same sparse MoE design with 2.8 trillion total parameters, 104 billion active per token, multimodal input, and a 1-million-token context window. Gains come from large-scale joint reinforcement learning across hundreds of downstream domains. Reported benchmark figures include Terminal-Bench 4.0 at 65.7%, ahead of GPT-6 Astra at 58.2% but behind Claude Opus 5.5 at 66.4%; Kimi Code Bench 3.0 Long-horizon at 61.2%, up from K3's 55.95%; BenchCAD at 92.7% with tools; and OfficeQA Pro v2 at 64.4. A faster K3.1 Pro Highspeed variant runs two to three times faster than the standard model.

OpenRouter's model page describes Kimi K3.1 Pro as Moonshot AI's flagship open-weight multimodal reasoning model and a post-trained successor to Kimi K3, released October 23, 2026 with a 1M-token context window. It targets long-horizon coding, 3D creation, financial research, and end-to-end agentic knowledge work, with native image and video understanding. The page lists the model identifier moonshotai/kimi-k3.1-pro with a headline rate of $0.376 per million input tokens and $10 per million output tokens, and notes that different companies host the same weights under Balanced, Nitro, or Floor routing. Because the page is dominated by hosting-provider pricing, latency, and uptime rows, it is used here only for model-level release facts.

NVIDIA's model card for Moonshot AI's Kimi K3.1 Pro documents it as a 3T-class multimodal model for reasoning, coding, and agentic workflows, accepting text, images, and video and generating text. The card lists a 1,048,576-token input context length, a Transformer Mixture-of-Experts architecture, and 2.8T total parameters with 104B active, based on Kimi K3. Developer-facing details include a 160,000-token vocabulary and RGB and MP4/WebM video input formats, with Moonshot recommending images up to 4096×2160 and video up to 1920×1080 as processing suggestions rather than hard limits. Documented use cases span code generation, image and video understanding, retrieval-augmented generation, document question answering, tool-assisted reasoning, CAD scripting, and financial research, with a Build.NVIDIA.com release date of 11/04/2026.

Moonshot AI introduced Kimi K3.1 Pro as the first model in its K3.1 series, with a 1-million-token context window and improvements in coding, 3D creation, financial research, and agentic work. The company also announced Kimi K3.1 Pro Highspeed, claimed to deliver two to three times the speed of the standard model. Both variants were made available through Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with full weights promised by November 4. K3.1 Pro was post-trained from Kimi K3 through joint reinforcement learning across hundreds of domains, using improvements to the RL framework, long-context inference, training efficiency, sandbox infrastructure, and rollout scheduling.

Moonshot AI’s official Hugging Face card describes Kimi K3.1 Pro as an open-weight, natively multimodal agentic model with 2.8 trillion total parameters, 104 billion activated per token, and a 1,048,576-token context window. It supports text and image input and uses MXFP4 weights with MXFP8 activations for quantization-aware training. The model is a Mixture-of-Experts architecture with 93 layers, 896 experts, and 16 experts selected per token. Relative to Kimi K3, it replaces full-attention Gated MLA layers with Gated DeepSeek Sparse Attention layers, while its 401M-parameter MoonViT-V2 encoder handles vision.