Tencent open-sources WeVisOne-VL: one VLM for detection, OCR, segmentation and matting
lmoroney · x · 2026-10-26
Tencent's WeChat Vision team released WeVisOne-VL (2B/4B, Apache 2.0, built on Qwen3-VL), unifying object-level and pixel-level perception in a single model: open-vocabulary detection, referring localization, counting, pointing, OCR, referring segmentation, and alpha matting.
- Boxes/points come out as coordinate tokens; a tiny 9.5M-parameter pixel decoder turns them into masks and full-resolution mattes
- Trained with three-stage curriculum SFT plus mixed-task RL; leading results across 30+ benchmarks (author-reported)
- Standout detail: it returns None when the requested object isn't in the image — reliable "not there" answers make pipelines safer
- Caveats: images only for now; fine edges still sensitive to resolution and occlusion
Suggested test: composite a dozen of your own photos, include a few where the target is missing, and count how often it answers honestly.
More from Models
- Users Report DeepSeek V4.1-Flash-1017 Update Regressed on Terminal-Bench 4.0 — teortaxesTex · 2026-10-26
- Users Report ChatGPT Suddenly Calling Them by Name in Every Reply — astralmatrix · 2026-10-26
- Theo fires back at Opus 5.5 critics with list of 8 projects shipped in a month — Vjeux · 2026-10-26
- Reddit asks: best low-refusal local model for RE, coding and red teaming on one RTX 5090 — Live_Drive_6256 · 2026-10-26
- DeepSeek 1017 checkpoint lands on FrontierCS 2 with a small score jump — scaling01 · 2026-10-26
- Developer: Claude models make 'crazy mistakes' after just two days away — ivan_bezdomny · 2026-10-26