https://t.co/1Po1IvW12o
AI Weekly's analysis
→
- dots studio, a Xiaohongshu subsidiary, released dots3-note preview: a 280B-parameter Mixture-of-Experts multimodal model with 16B active parameters, under Apache 2.0.
- The model handles text, images, video with audio tracks, and audio transcription in a single 512K-token context using an MoE ViT and 800M audio encoder.
- Self-reported benchmarks include 78.4% on SWE-bench Verified, 79.1% on MMMU Pro, and 52.6% on HLE; BF16 and FP8 checkpoints ship for SGLang and vLLM.
Read full analysis →