Qwen-VL (Patch Grid)
qwen_vl_utils.smart_resize). The token clamp defaults shown here are the library defaults; a DashScope endpoint may cap lower via a per-request max_pixels. Verify against your deployment.How Qwen-VL bills images
Alibaba Qwen2-VL / Qwen2.5-VL use a 14px ViT patch with a 2×2 spatial merge, giving a 28px effective grid (image_factor = 14 × 2 = 28). smart_resize rounds each side to a multiple of 28; the image token count is the number of merged patches:
tokens = (W/28) · (H/28)
bounded by [IMAGE_MIN_TOKEN_NUM, IMAGE_MAX_TOKEN_NUM] = [4, 16384] tokens (the qwen_vl_utils defaults: max_pixels = 16384 × 28²). Production DashScope tiers often set a lower max_pixels per request.
Does snapping help?
The patch is small (28px), so spill-over per step is modest — but two effects still pay off:
- Snapping to the 28px grid removes fractional rows/columns that round up.
- Padding strip lowers the pixel area, which directly lowers the patch count until the
max_pixelsclamp is hit.
For high-resolution inputs the image is first scaled under max_pixels, so the win is often capped — Qwen's value is more about file size + context budget than dramatic token cuts.
vision-squeezer image.png --model qwen
CLI aliases: qwen, qwen-vl. MCP target_model: "qwen".
Token savings
The 28px patch is small, so the lever here is area (padding strip), not boundary snapping.
| Scenario | Before | After | Saved |
|---|---|---|---|
| 1024×1024 → strip solid border to 896×896 | 1,369 tok | 1,024 tok | −25% |
| 1024×1024 → 28px grid snap only (1008×1008) | 1,369 tok | 1,296 tok | −5% |
Cropping wasted area pays off; pure grid snapping is incremental. For very large inputs the count is also capped at 16,384 tokens, so savings there come from staying under that ceiling.
Source
Grid and clamp taken from the official qwen_vl_utils.smart_resize reference implementation (QwenLM/Qwen2.5-VL) — image_factor = patch_size(14) × spatial_merge(2) = 28, bounds [IMAGE_MIN_TOKEN_NUM, IMAGE_MAX_TOKEN_NUM] = [4, 16384]. The same 28px grid carries across Qwen2-VL, Qwen2.5-VL, and Qwen3-VL (identical vision_process.py). Background: Qwen2.5-VL technical report (arXiv:2502.13923, Feb 2025). Verified 2026-06-11.