Moonshot AI has released Kimi K3.1 Pro, its most capable 3T-class model to date. As an open-weight, natively multimodal agentic model, it is designed to handle both text and image inputs. The model was post-trained from Kimi K3 using large-scale joint reinforcement learning across hundreds of downstream domains. It features a 1-million-token context window and aims to deliver significant improvements in coding, 3D creation, financial research, and factual accuracy.
This image illustrates the conversational interface of the Kimi AI model family.
The model utilizes a Mixture-of-Experts (MoE) architecture, totaling 2.8 trillion parameters with 104 billion activated parameters per token. A key architectural shift from the previous Kimi K3 model is the replacement of full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA) layers. This change is specifically intended to facilitate more efficient long-context modeling. The architecture includes 896 experts, with 16 experts selected per token and 2 shared experts, alongside a vision encoder (MoonViT-V2) with 401 million parameters.
In performance evaluations, Kimi K3.1 Pro shows competitive results across various benchmarks. For example, it achieved a score of 65.7 on Terminal-Bench 4.0 and 92.7 on BenchCAD with tools, placing it alongside other high-performing models like Claude and GPT-6 Astra in several categories. However, the source notes that some results, such as those for the PostTrain Bench, were evaluated using H20 GPUs rather than H100 GPUs, which may affect direct comparisons with models tested on higher-end hardware. Additionally, for multi-turn conversations, the model requires the complete assistant message—including the `reasoning_content` field—to be passed back to the API to maintain proper context.
While the model is open-weight, its massive scale presents significant deployment considerations. The architecture's complexity and parameter count suggest high hardware requirements for local inference. Furthermore, while the model performs well on internal benchmarks like Kimi Code Bench 3.0, these results are specific to the developer's internal testing environment and may vary in different production settings. The model is currently available via API and supported on inference engines such as vLLM, SGLang, and TokenSpeed.