Kimi K3.1 Pro weights released with sparse attention improvements
Moonshot AI has released open-weight model files for Kimi K3.1 Pro, a multimodal agentic model that represents the first in the K3.1 series. The model builds on Kimi K3 through large-scale joint reinforcement learning across hundreds of downstream domains and supports a 1-million-token context window.
Architecturally, the model replaces full-attention Gated MLA layers with Gated DeepSeek Sparse Attention layers to enable more efficient long-context modeling. The release delivers improvements in coding, 3D creation, financial research, factual accuracy, and agentic capabilities relative to its predecessor.
The weights are available through multiple inference routes including Transformers, vLLM, SGLang, and Docker, along with hosted options via HuggingChat and notebook environments such as Google Colab and Kaggle. Comprehensive evaluation results across coding and agentic task benchmarks demonstrate performance across diverse downstream domains.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.