Tencent's WeChat Vision team open-sources WeVisOne-VL 2B and 4B under Apache 2.0: one model for detection, grounding, counting, OCR, segmentation and matting
Perception tasks that used to need several vision models are folded into one small model. This covers what it does and where it stops.
// Key points
- Built on Qwen3-VL, it casts object-level tasks (open-vocabulary detection, referring localization, counting, pointing, OCR) and pixel-level tasks (referring segmentation, promptable alpha matting) as structured next-token prediction.
- Pixel-level output comes from a lightweight decoder of about 9.6 million parameters; boxes and points use coordinate tokens from <0> to <1000>, and the model returns None when a target is absent.
- For inference efficiency it emits structured output directly without chain-of-thought, and it currently handles single-turn images only, not video.
- Tencent-authored code is released under Apache 2.0, with third-party components keeping their own licenses; it is positioned for research and not for autonomous high-impact decisions or identifying individuals.
Builder's takeSearch-by-content and locating text in screenshots in an AI cloud drive still means chaining several models today, so this 4B model is worth a consolidation trial. I'd compare OCR localization and counting accuracy on my own image samples first; it doesn't handle video and explicitly shouldn't be used to identify people, so cross both off in the product design up front.
// Background · from #Open models
Full timeline →- Oct 25 Hugging Face launches a Trainium compatibility lookup: whether a Hub model runs unmodified on Trainium, with a script to run it yourself
- Oct 24 AWS shows the full bill for one multi-turn RL run: about $339 took gpt-oss-20b from a 61% to an 84% Wordle solve rate
- Oct 23 SageMaker AI training sessions are generally available: write your own post-training loop and reward function while AWS runs the GPUs, billed by tokens processed