Qwen's VisionWeave saves 43% visual tokens at 98.9% performance, 2.3x throughput
Qwen · hf · 2026-10-12
Core idea
Alibaba's Qwen team introduces VisionWeave, establishing "elastic visual representation weaving" as a native MLLM capability: the model learns end-to-end where — and at what granularity — to allocate visual representations, via two components:
- Gated spatial pooler: builds coarse-grained representations alongside native fine-grained ones in a shared MRoPE coordinate
- Granularity router: learns content-adaptive allocation between them
Training and results
- Validated on Qwen3.5-4B via self-distillation alone, scaled to Qwen3.8-27B with 30K+ A100 GPU-hours
- Saves 43.0% tokens on average while retaining 98.9% native performance across eight benchmarks; token-pruning baselines at a fixed 50% savings preserve only 88%
- On SGLang serving: 2.3x throughput, 54.4% lower mean TTFT, 60.6% lower mean TPOT
More from Infra
- Strix Halo llama.cpp branch boosts prompt processing 21–54% over Vulkan — asfbrz96 · 2026-10-12
- Chinese chipmaker CXMT hits design breakthrough in tech sought by Samsung, SK Hynix — AIFlow_ML · 2026-10-12
- Same model, different behavior: Reddit asks how to test open-model inference compatibility — Top_Speaker_7785 · 2026-10-12
- The memory price crisis is far worse than you think, video explains — Brassens71 · 2026-10-12
- Dev Picks Up AMD Halo Mini PC, Praises Its Heat Management for Local AI — gajesh · 2026-10-12
- SoloQ: calibration-free 4-bit quantization cuts dLLM memory 2.61x, speeds inference 2.24x — UniversityofSouthernCalifornia · 2026-10-12