{"slug": "vlm-run-puts-six-open-ocr-models-behind-one-openai-compatible-endpoint", "title": "VLM Run puts six open OCR models behind one OpenAI-compatible endpoint", "summary": "VLM Run, founded by CEO Sudeep Pillai, launched an OpenAI-compatible gateway on August 19 that routes document-processing work across six open OCR and vision-language models, including GLM-OCR, DeepSeek-OCR-2, dots.mocr, PP-OCRv6, PaddleOCR-VL, and Florence-2-base-ft, without requiring developers to rewrite client integrations. The gateway handles rasterization, page splitting, order preservation, streaming, retries, and memory management, aiming to insulate developers from rapid model churn. Pillai, who holds a robotics PhD from MIT and led ML work at Toyota Research Institute, positions the API layer as the durable part of document AI, with the service also exposed via an MCP server for agent integration.", "body_md": "# VLM Run puts six open OCR models behind one OpenAI-compatible endpoint\n\n**Founder Sudeep Pillai is betting orchestration will matter longer than any single document model.**\n\nBy [RuntimeWire Staff](/author/runtimewire-staff)\n· Published\n\nPrimary source: [Hugging Face Newsroom](https://huggingface.co/blog/vlm-run/intro-to-vlmrun-gateway)\n\n## Why it matters\n\nVLM Run is selling insulation from rapid model churn. If its cost and accuracy claims hold across real documents, developers can switch OCR models without rebuilding production pipelines.\n\n[Sudeep Pillai](https://www.vlm.run/about?ref=runtimewire), founder and CEO of [VLM Run](https://vlm.run/?ref=runtimewire), launched an OpenAI-compatible gateway on August 19 that lets developers route document-processing work across open-weight OCR and vision-language models without rewriting their client integration.\n\nThe [VLM Run Gateway](https://www.vlm.run/product/gateway?ref=runtimewire) initially supports OCR-focused models including GLM-OCR, DeepSeek-OCR-2, dots.mocr, PP-OCRv6, PaddleOCR-VL and Florence-2-base-ft. Developers point the OpenAI SDK at VLM Run's endpoint, select a model and send a document URL through the familiar chat completions interface, according to [VLM Run's launch article on Hugging Face](https://huggingface.co/blog/vlm-run/intro-to-vlmrun-gateway?ref=runtimewire).\n\nPillai's bet comes from a career spent making perception systems work outside research demos. He earned a robotics PhD at MIT and later led machine-learning work at Toyota Research Institute, according to [VLM Run's team page](https://www.vlm.run/about?ref=runtimewire). VLM Run's founding researcher, [Dinesh Narapureddy](https://www.ri.cmu.edu/alumni/dinesh-reddy-narapureddy/?ref=runtimewire), completed his robotics PhD at Carnegie Mellon University in 2023 with a thesis on self-supervising occlusions for computer vision. The rest of VLM Run's early technical group includes former AWS staff and computer-vision researchers from North Carolina State University.\n\nThat background explains the problem VLM Run chose to tackle. Model evaluation is only one part of document AI. Production systems also have to rasterize PDFs, split and distribute pages, preserve page order, stream results, retry failed jobs and contain out-of-memory errors. VLM Run Gateway handles those operations behind the endpoint, according to VLM Run, so a developer can change the underlying model without rebuilding the surrounding document pipeline.\n\n### An escape hatch from model churn\n\nVLM Run is selling insulation from a market where the preferred document model can change within months. A team may choose a model after a short evaluation, then spend far longer building batching, parsing, monitoring and retry logic around its API. A better model can arrive before that integration work has paid for itself.\n\nGateway makes the API layer the durable part. The [model catalog](https://docs.vlm.run/gateway/models?ref=runtimewire) spans systems with sharply different sizes and intended uses, from the 22 million-parameter PP-OCRv6 to the 3 billion-parameter dots.mocr and DeepSeek-OCR-2. Developers can use JSON mode and typed output contracts, while each response includes a `usage.cost`\n\nfield intended to make side-by-side testing easier.\n\nVLM Run has also exposed the service through an [MCP server](https://docs.vlm.run/gateway/mcp-server?ref=runtimewire). Its `read_document`\n\ntool allows compatible coding agents and clients to submit long documents through the same processing layer. That gives VLM Run a second distribution path: developers can integrate Gateway directly into applications, or let agent software call it as a tool.\n\nThe approach puts VLM Run between model developers and the applications consuming their work. Pillai does not need one OCR model to dominate. Frequent model releases strengthen the case for a neutral routing and orchestration layer, provided VLM Run can keep integrations current and prove that switching models does not introduce unacceptable output differences.\n\n### The price claim needs the accuracy data\n\nVLM Run says Gateway can process more than 100,000 pages for under $60 and can be around 10 times cheaper than frontier vision models for extraction, OCR, layout recognition and parsing. Those figures remain VLM Run's estimates, and the launch article does not include the document set, token assumptions, latency measurements or accuracy scores behind them.\n\nVLM Run has not published Gateway-specific pricing or the methodology behind the under-$60 estimate in the supplied materials. The company's broader [pricing page](https://www.vlm.run/pricing?ref=runtimewire) offers a $10 signup balance and a $799 monthly Pro plan with $1,000 in included usage. Enterprise options include private deployments, custom rate limits and compliance features.\n\nAccuracy will decide whether those savings survive contact with production documents. A cheap parser that loses a signature, reads an invoice total incorrectly or scrambles a dense table can create costs elsewhere in the workflow. VLM Run says the strongest model varies by language, layout density, scan quality and domain, which is a sound argument for routing. Gateway's launch leaves customers to run those comparisons themselves while VLM Run prepares a public version of its internal accuracy leaderboard.\n\n### A focused entry into visual infrastructure\n\nGateway is a narrower product than VLM Run's Orion visual agent, which handles images, documents and video through a mixture of vision-language models and specialized computer-vision tools. Documents give Pillai an immediate entry point because the task is bounded, the costs are measurable and buyers already understand the operational pain.\n\nThe competitive field includes Amazon Textract, Microsoft Azure Document Intelligence and newer document-processing platforms such as Reducto, LlamaParse and Extend. VLM Run's chosen distinction is model choice through an OpenAI-compatible interface. That positioning also avoids asking developers to adopt another proprietary SDK before they can test the service.\n\nVLM Run was founded in 2023 and operates in the San Francisco Bay Area, according to [South Park Commons](https://www.southparkcommons.com/companies/vlm-run/?ref=runtimewire), where Pillai is listed as a Founder Fellow. VLM Run identifies South Park Commons, Gallery and York IE as backers, though it has not attached round sizes or a valuation to those names.\n\nPillai is building around a simple expectation: visual models will keep improving faster than production teams can replace their infrastructure. Gateway turns that churn into the product. VLM Run's next test is whether its routing layer can produce enough verified accuracy and cost data to become the default place developers compare the models beneath it.", "url": "https://wpnews.pro/news/vlm-run-puts-six-open-ocr-models-behind-one-openai-compatible-endpoint", "canonical_source": "https://runtimewire.com/article/vlm-run-gateway-open-weight-ocr-models", "published_at": "2026-08-21 10:42:52+00:00", "updated_at": "2026-08-21 11:16:58.306183+00:00", "lang": "en", "topics": ["ai-products", "ai-infrastructure", "developer-tools", "computer-vision", "natural-language-processing"], "entities": ["VLM Run", "Sudeep Pillai", "Hugging Face", "GLM-OCR", "DeepSeek-OCR-2", "dots.mocr", "PP-OCRv6", "PaddleOCR-VL"], "alternates": {"html": "https://wpnews.pro/news/vlm-run-puts-six-open-ocr-models-behind-one-openai-compatible-endpoint", "markdown": "https://wpnews.pro/news/vlm-run-puts-six-open-ocr-models-behind-one-openai-compatible-endpoint.md", "text": "https://wpnews.pro/news/vlm-run-puts-six-open-ocr-models-behind-one-openai-compatible-endpoint.txt", "jsonld": "https://wpnews.pro/news/vlm-run-puts-six-open-ocr-models-behind-one-openai-compatible-endpoint.jsonld"}}