The Pulse
Moonshot Makes Kimi K3.1 Pro’s Open Weights Available
Kimi K3.1 Pro is a 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window. Moonshot lists API access and currently recommends vLLM, SGLang and TokenSpeed for inference.
AI.info Team · Written with AI · fact-checked ·
Moonshot’s first K3.1 model
The model card describes Kimi K3.1 Pro as an open-weight, natively multimodal agentic model and the first model in Moonshot’s K3.1 series. Moonshot says it was post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of downstream domains. The model card describes it as the company’s most capable 3T-class model to date, and says it delivers substantial improvements in coding, 3D creation, financial research, factual accuracy and other agentic capabilities. Its listed modalities are text and image.
Inside the 2.8T-parameter model
The model uses a mixture-of-experts architecture, with 2.8 trillion total parameters and 104 billion activated parameters. Its 93 layers include one dense layer, and each token selects 16 experts from a pool of 896, alongside two shared experts. The attention system combines 69 KDA layers with 24 Gated DSA layers; Moonshot says that, compared with Kimi K3, it replaced the full-attention Gated MLA layers with Gated DSA layers for more efficient long-context modeling. The model card lists 96 attention heads and an attention hidden dimension of 7168.
Kimi K3.1 Pro has a 1-million-token context window. The vision encoder is MoonViT-V2, with 401 million parameters. The card specifies MXFP4 weights and MXFP8 activations, using quantization-aware training.
Scores span coding, agents and reasoning
Moonshot’s evaluation table reports Kimi K3.1 Pro results at maximum reasoning effort, temperature 1.0 and top-p 0.95 unless otherwise specified. In coding, it records 65.7 on Terminal-Bench 4.0, 38.1 on Terminal-Bench-Science 0.1, 51.7 on SWE Marathon v1.1 and 31.5 on Program Bench. The card says Program Bench uses a binary score: a task receives 1 when its ProgramBench score reaches at least 95%, and 0 otherwise.
For agentic tasks, the table gives the model 70.0 on HLE-Full with tools, 92.1 on BrowseComp and 43.2 on Agents’ Last Exam (ALL). On reasoning and knowledge benchmarks, it reports 94.2 on GPQA-Diamond, 51.1 on HLE-Full and 39.4 on HLE-Diamond. The card says the full HLE evaluation sets contain 2,500 and 1,000 questions, respectively.
The cybersecurity results include 74.7 on ExploitBench and 20.7 on ExploitGym. The ExploitBench score uses the official harness with a 300-turn budget and reports the pass@5 maximum-capability score. For ExploitGym, Kimi K3.1 Pro is evaluated with Kimi Code under a six-hour wall-clock budget per task.
API access, inference engines and license
Moonshot makes the model’s API available through platform.kimi.ai by selecting kimi-k3.1-pro, and says it provides an OpenAI/Anthropic-compatible API. The model card currently recommends vLLM, SGLang and TokenSpeed as inference engines.
Reasoning effort is set with the reasoning_effort request field, which supports low, high and max, with max as the default. For multi-turn conversations and tool calls, the card says the complete assistant message returned by the API must be passed back to messages as-is, including reasoning_content and tool_calls, not just content. Kimi K3.1 Pro is released under the Kimi K3.1 Pro License.