Xiaohongshu opens dots3-note, a 280B MoE multimodal model

Xiaohongshu's dots studio has posted its first open-weight release in a new model family. The model card on Hugging Face describes dots3-note preview as a Mixture-of-Experts multimodal system with 280 billion total parameters and 16 billion active during inference, published under Apache 2.0.

Architecture is where the choices show. The model uses 256 routed experts plus one shared expert with top-8 routing, arranged across one dense and 45 MoE layers, with a 5,120 hidden size and a 152K vocabulary. Context runs to 512K tokens. A separate MoE vision encoder (7B total, 1.2B activated) and an 800M dense audio encoder let the same model handle text, images, video with audio tracks, and audio transcription.

For benchmarks, dots studio reports 78.4% on SWE-bench Verified, 61% on SWE-bench Pro, 75.7% on SWE-bench Multilingual, 79.1% on MMMU Pro, and 52.6% on HLE, along with Claw-Eval and Apex Agents scores. The card pitches tool use and multi-step agent workflows as first-class capabilities. Two checkpoints ship, a BF16 full-precision version and an FP8 quantised one recommended for deployment, with launch commands provided for SGLang and vLLM.

These are self-reported preview numbers on a fresh release. The card does not compare against named peers, does not describe the training corpus or compute budget, and says nothing about safety filtering or provenance for the multimodal data. For teams that want a permissively licensed, self-hostable multimodal model with a 512K-token context, another serious Chinese-lab entry is now on the shelf, and a free tier on OpenRouter lets anyone poke at it before committing GPUs.