VLM Run Puts Open-Weight OCR Models Behind an OpenAI-Compatible Gateway

VLM Run has introduced Gateway, a single OpenAI-compatible endpoint for running open-weight OCR and vision-language models on documents. The service supports models including GLM-OCR, DeepSeek-OCR-2, dots.mocr, PaddleOCR-VL, PP-OCRv6 and Florence-2-base-ft.

The pitch is simpler model experimentation without rebuilding an entire document pipeline. Developers can submit a PDF URL, while Gateway handles page rasterization, worker distribution, page ordering, streaming, retries and out-of-memory isolation. That abstraction is designed to make it easier to switch models as accuracy and cost requirements change.

Built for document workflows

Gateway also supports JSON mode with typed output contracts, making it suitable for structured extraction. An MCP server provides a read_document tool for compatible agent clients such as Claude Code, Codex and OpenCode, allowing those systems to process visual documents through the same service.

VLM Run says open-weight OCR models can be substantially less expensive than frontier VLMs for extraction, layout analysis and parsing. Its article claims that processing more than 100,000 pages can cost under $60 in some scenarios, with production savings often reaching roughly 10x. Each response includes usage.cost, enabling teams to compare expenses alongside output quality.

The catalog currently spans models from 0.2B to 3B parameters, and the company says additional visual and audio models are also hosted. A CLI, Python client, Colab quickstart and MCP documentation are available for testing.

The main caveat is that no single OCR model is presented as universally best. Healthcare forms, drawings, contracts, languages and scan quality can produce different results, so developers will still need evaluations on their own data. VLM Run says a public accuracy leaderboard is planned; for now, the Gateway is positioned as a faster way to run those comparisons.

Source: Hugging Face Blog