Moonshot AI has released Kimi K3.1 Pro, an open-weight, natively multimodal agentic model with 2.8 trillion total parameters and 104 billion activated parameters. Built on a Mixture-of-Experts architecture with 93 layers, it features a 1-million-token context window and advanced attention mechanisms including KDA and Gated DSA. The model supports text and image modalities, powered by a Moon ViT-V2 vision encoder with 401M parameters, and uses MXFP4 quantization for efficiency.
Post-trained from Kimi K3 with large-scale joint reinforcement learning across hundreds of domains, Kimi K3.1 Pro shows strong performance in coding, 3D creation, financial research, and factual accuracy. It outperforms or matches leading models like Claude Fable 5.1, GPT-6 Astra, and DeepSeek-V4.1 across multiple benchmarks including Terminal-Bench, SWE Marathon, and Browse Comp. Notably, it scores 92.1 on Browse Comp and 95.4 on Baby Vision, demonstrating frontier-level agentic and multimodal capabilities.
Available on Hugging Face, Kimi K3.1 Pro represents a significant leap in open-weight model performance, combining massive scale with specialized training for real-world agent tasks.
Source: Hacker News · Summarized by HeadlinesBriefing