280B total / 16B active (multimodal MoE)
524,288 tokens
Apache License 2.0
First open-weight model in Xiaohongshu's dots3 family and its most lightweight member, designed as a natively multimodal general-purpose and agentic model accepting text, images, video, and audio while producing text. Its language backbone has one dense layer followed by 45 MoE layers, a 5,120-dimensional hidden state, 256 routed experts plus one shared expert with top-8 routing, 13 DSA attention layers, 33 sliding-window-attention layers, and a 152K vocabulary. The multimodal stack adds a sparse MoE vision transformer with 7B total and 1.2B activated parameters plus a dense 800M-parameter audio encoder. Official positioning emphasizes mathematical and logical reasoning, code generation, tool use, multi-step agent workflows, long-context processing, and interactive tasks involving exploration, memory updates, and adaptation. The chat template supports switchable thinking through enable_thinking, while native structured tool calls are supported through the dots tool-call format. Published headline evaluations report 73.4 on Claw-Eval, 61.0 on SWE-bench Pro, 83.3 on BrowseComp, 81.4 on ARC-AGI-2, 75.1 on Terminal-Bench 2.1, and a 3056 Codeforces rating. Multimodal claims include 83.1 on CharXiv-RQ, 72.5 on SimpleVQA, 60.7 on GDPval PDF, 39.3 on MMEVideo v2, and 69.8 on MMAU-Pro. BF16 and FP8 checkpoints are provided, with the FP8 release recommended for one eight-GPU node. Native vLLM support is available on main, while Transformers and SGLang integration pull requests remain under review. The architecture includes a shared 1.13B-parameter Multi-Token Prediction layer; optional MTP/NEXTN speculative decoding is claimed to reduce time per output token by more than 50 percent.