Community impressions of OvisOCR2, a 0.8B local document parser
A new 0.8B end-to-end OCR model called OvisOCR2, based on Qwen3.5-0.8B, converts full document pages into structured Markdown. It scores 96.58 on OmniDocBench v1.6 and 75.06 on PureDocBench, with weights available under Apache 2.0 and vLLM support. The community notes its impressive performance for the size but calls for independent real-world testing.
Entities
Related
Local Context Router cuts vision-token waste by pre-classifying PDF pages
A developer built LocalContextRouter, a tool that runs locally to classify each PDF page as text, OCR-needed, or image before sending to a multimodal model. This avoids the wasteful practice of rendering every page to an image, which can cost 1,300–4,800 tokens per page versus 400–800 tokens for plain text. The tool never calls a model itself, leaving the final API call to the user's app.
Open-source 27B model on Groq achieves 85% on ExtractBench, frontier models score 0%
A 27B open-source model running on Groq scored 85% on a 369-field extraction task from SEC filings, while six frontier models scored 0%. The benchmark, ExtractBench (arXiv:2602.12247), uses real documents with human-checked answers.
LocalEyes gives blind LLMs vision via local Ollama models
LocalEyes is a new tool that enables text-only LLMs like DeepSeek, CodeLlama, and Qwen-Coder to process images locally using an Ollama vision model. It supports screen capture, clipboard reading, and image file analysis without cloud uploads or API keys, offering a private, fast, and free solution for developers using Claude Code.
VisionBridge: OpenAI-compatible proxy gives text-only LLMs vision via separate vision models
VisionBridge is a tiny OpenAI-compatible proxy that enables text-only reasoning models (DeepSeek, Qwen, GLM) to process images by querying a separate vision model through tool calls like look, OCR, scan, crop, and compare. It requires no training or weight modifications, working with LM Studio, Ollama, vLLM, and other OpenAI-compatible backends.
OmniMapBench benchmark tests LVLMs on visual-centric map reasoning
Researchers introduced OmniMapBench, a benchmark of 2,096 QA pairs across 1,603 map documents from nine categories, designed to evaluate visual-centric reasoning in large vision-language models. It addresses the limitation that many document benchmarks allow high performance via text-only cues, requiring genuine visual grounding for tasks from perception to multi-step reasoning.