191–194 of 194 posts
Ironically, the pill counting example selected to showcase "the best vision model" can be easily solved with OpenCV template matching, a technology created 25 years ago.
The point is that it's general. It can do this task and many other tasks and it doesn't need custom development like OpenCV does. Of course if you only want to count pills and you want it to be cheap/fast you're still better off using OpenCV.
That said, considering variables like cost (say, over 100k PDF pages) and accuracy requirements (e.g., construction documents with dense images & tables), Gemini and other SOTA VLMs are expensive and inaccurate, and therefore unsuitable. This is where niche, open-weight, and task-specific OCR/VL models come in.
With a one-line change, you can switch between DeepSeek OCR 2, GLM-OCR, dots.mocr, Paddle OCR VL, PP-OCRv6, etc., and process 100K+ pages for under $60 on VLM Run Gateway
This is why we built VLM Run Gateway, one OpenAI-compatible endpoint for open-weight OCR and VLM models.
Try it out quickly via OpenAI SDK:
``` client = OpenAI( base_url="https://gateway.vlm.run/v1/openai", api_key="", )
response = client.chat.completions.create( model="rednote-hilab/dots.mocr", messages=[{ "role": "user", "content": [{ "type": "document_url", "document_url": {"url": "https://.../invoice.pdf"}, }], }], extra_body={"document_dpi": 72}, ) ```
or via our CLI:
``` pip install vlmrun vlmrun gw models vlmrun config set --api-key 'vlmrun' # anon-user, rate-limited vlmrun gw chat .pdf -m zai-org/glm-ocr vlmrun gw chat .pdf -m zai-org/glm-ocr --json-mode vlmrun gw chat .pdf -m deepseek-ai/deepseek-ocr-2 vlmrun gw chat .pdf -m rednote-hilab/dots.mocr vlmrun gw chat .pdf -m paddleocr/pp-ocrv6 ```
Docs: https://docs.vlm.run/gateway
Catalog: https://docs.vlm.run/gateway/models
MCP: https://docs.vlm.run/gateway/mcp-server
Colab Quickstart: https://colab.research.google.com/drive/1RkuVIyuc5Po-UlcSlFy...
Read the full post here: https://huggingface.co/blog/vlm-run/intro-to-vlmrun-gateway
I actually favor Qwen3.8 and run it locally + use the Token-Plan on AlibabaCloud, when I need faster results. Kind of favor it over GPT5.6 Sol. Also it seems to be more capable, need to test more, but I think it's at least getting on par and it's fully open-source and open-weights. Here's some benchmarks: https://benchlm.ai/compare/gpt-5-6-sol-vs-qwen3-8-max https://qwen.ai/blog?id=qwen3.8#full-benchmark-table (incre…
I thought Qwen 3.8 max doesn't have vision?
You can get this to run on 24GB Ram: https://huggingface.co/baa-ai/Qwen3.8-27B-RAM-24GB-MLX
I found the FULL Qwen3.8-2.4T-A95B-MLX-reap50-3bit, but it's 540GB. So, I can't run it on my tiny laptop, but hope that the community finds ways to bring the size down and memory requirements too =)))
Earlier quoted context omitted.
I thought Qwen 3.8 max doesn't have vision?
It actually has vision + tool_use even the 27B param model. The community tries to produce even a MoE version of Qwen3.8 now, because that'd allow to run the full model with some experts being pruned like with Ornith 1.5 35B A3B. You can get this to run on 24GB Ram: https://huggingface.co/baa-ai/Qwen3.8-27B-RAM-24GB-MLX I found the FULL Qwen3.8-2.4T-A95B-MLX-reap50-3bit, but it's 540GB. So, I can't run it on my tiny…
Research preview: Whittle MoE 27B (A18B): a mixture of experts rescued by its routers
https://huggingface.co/logic65/Qwen3.8-Whittle-MoE-27B-A17.8...