Releases·Tool Release
Skill issue: stop deploying vision language models, use them with Skills — Merve Noyan, Hugging Face
In a talk, Hugging Face's Merve Noyan argues developers should not call vision language models at inference time, since task-trained small detectors are faster and more accurate. She demos a toolkit that gives coding agents Apache 2.0 CV models and a 'vibe training' pipeline that auto-labels data and trains RF-DETR for a few dollars.
61 engagement1 sourceWed, Sep 23, 2026, 01:30 PM
AI Overview
- Merve Noyan's toolkit has a coding agent label unlabeled data with a 9B-parameter open VLM, verify with two VLM judges under minimum agreement, and train RF-DETR for $3–4 on Hugging Face jobs
- Trained detector hits good mAP on road signs and generalizes to document parsing, catching a signature the labeling VLM missed
- Findings: judges reject at unequal rates so consensus starves data, judge prompts need human approval, coding agents flip traffic signs and jitter traffic-light colors
- Tools half spans Apache 2.0 picks from a 600M-parameter segmentation model to pose, depth, and OCR, with image-guided detection planned
Related
Products·launched
OpenAI-compatible API gateway for serving open-weight VLMs, OCR models and vision models
A gateway that exposes open-weight vision-language models, OCR VLMs, and ViT-based vision models behind a single OpenAI-compatible API. It addresses production problems with serving VLMs, such as providers serving different quantizations under the same model id, which harms visual accuracy. Aimed at developers running these models in production.
A gateway that exposes open-weight vision-language models, OCR VLMs, and ViT-based vision models behind a single OpenAI-compatible API. It addresses production problems with serving VLMs, such as providers serving different quantizations under the same model id, which harms visual accuracy. Aimed at developers running these models in production.
5 engagement1 sourcehackernewsFri, Sep 4, 2026, 06:31 PM
Products·building
Multimodal computer vision agent where a VLM prompts a segmentation model for object tracking
A demo system where a vision-language model acts as the brain and prompts a fast segmentation model (SAM 2.1 Tiny) to track objects, e.g. in a cup and ball game. It avoids VLM context-window limits by delegating per-frame segmentation to a task-specific model. Built as a side project demo.
A demo system where a vision-language model acts as the brain and prompts a fast segmentation model (SAM 2.1 Tiny) to track objects, e.g. in a cup and ball game. It avoids VLM context-window limits by delegating per-frame segmentation to a task-specific model. Built as a side project demo.
1 engagement1 sourceredditFri, Sep 4, 2026, 08:01 AM
Releases·Tool Release
VisionBridge and LocalEyes: tools give text-only LLMs vision via separate models
Two open-source tools, VisionBridge and LocalEyes, enable text-only reasoning models (DeepSeek, Qwen, GLM, CodeLlama) to process images by querying a separate vision model. VisionBridge is an OpenAI-compatible proxy that uses tool calls (look, OCR, scan, crop, compare) and works with LM Studio, Ollama, vLLM, and others. LocalEyes integrates with Claude Code and uses a local Ollama vision model for screen capture, clipboard reading, and image analysis, all without cloud uploads. Both require no training or weight modifications.
Two open-source tools, VisionBridge and LocalEyes, enable text-only reasoning models (DeepSeek, Qwen, GLM, CodeLlama) to process images by querying a separate vision model. VisionBridge is an OpenAI-compatible proxy that uses tool calls (look, OCR, scan, crop, compare) and works with LM Studio, Ollama, vLLM, and others. LocalEyes integrates with Claude Code and uses a local Ollama vision model for screen capture, clipboard reading, and image analysis, all without cloud uploads. Both require no training or weight modifications.
Updated Tue, Jul 7, 2026, 07:13 PM — LocalEyes released on 2026-07-06, VisionBridge on 2026-07-07, both providing vision for text-only LLMs via separate visi
58 engagement2 sourcesgithubMon, Jul 6, 2026, 09:30 PM
Social·Community
GLM 5.3 Flash beat DeepSeek V4 Flash Vision Exp in a browser screenshot QA test
A developer running a browser QA harness that converts screenshots into actions compared DeepSeek V4 Flash Vision Exp and GLM 5.3 Flash, citing cost concerns with prior vision models at high screenshot volumes. GLM performed better on their test set. Both models ship under the MIT license with hosted inference options.
A developer running a browser QA harness that converts screenshots into actions compared DeepSeek V4 Flash Vision Exp and GLM 5.3 Flash, citing cost concerns with prior vision models at high screenshot volumes. GLM performed better on their test set. Both models ship under the MIT license with hosted inference options.
8 engagement1 sourceredditFri, Sep 4, 2026, 09:51 AM
Releases·Model Release
SenseNova-Vision: 7B open model unifies vision tasks without task-specific heads
SenseNova-Vision is a 7B mixture-of-transformers model released under Apache 2.0 that treats all computer vision tasks as a single generation problem, handling segmentation, depth, detection, OCR, and 3D reconstruction without task-specific heads. It accepts natural language instructions and visual hints, outputting text or images, simplifying vision pipelines for practitioners.
SenseNova-Vision is a 7B mixture-of-transformers model released under Apache 2.0 that treats all computer vision tasks as a single generation problem, handling segmentation, depth, detection, OCR, and 3D reconstruction without task-specific heads. It accepts natural language instructions and visual hints, outputting text or images, simplifying vision pipelines for practitioners.
15 engagement1 sourceredditThu, Aug 13, 2026, 03:07 PM