Six OCR models, one endpoint: VLM Run bets the API outlives the model

On August 19, VLM Run founder and CEO Sudeep Pillai launched an OpenAI-compatible gateway that routes document-processing work across open-weight OCR and vision-language models without changes to the client integration ^[1]. The interesting part is not the model list but the premise: the layer worth engineering carefully is the interface, because the model underneath it will be replaced.

The initial catalog is six models: GLM-OCR, DeepSeek-OCR-2, dots.mocr, PP-OCRv6, PaddleOCR-VL and Florence-2-base-ft ^[2]. According to VLM Run's launch article on Hugging Face, developers point the OpenAI SDK at the company's endpoint, pick a model and pass a document URL through the chat completions interface ^[3]. The models are not interchangeable in any ordinary sense: the catalog runs from the 22 million-parameter PP-OCRv6 to the 3 billion-parameter dots.mocr and DeepSeek-OCR-2 ^[4], roughly a 136-fold span in parameter count ^[5]. Presenting that range behind one call signature is the product.

The unglamorous work is the argument for it. Production document systems have to rasterize PDFs, split and distribute pages, preserve page order, stream results, retry failed jobs and contain out-of-memory errors; VLM Run says Gateway handles those operations behind the endpoint ^[6]. That is the expensive part of a document pipeline, and it is the part a team keeps rebuilding when it changes models. VLM Run's framing is that a team evaluates a model quickly, then spends far longer building batching, parsing, monitoring and retry logic around it, and a better model can land before that work has paid for itself ^[7]. Supporting features follow the same logic: JSON mode, typed output contracts, and a usage.cost field on every response intended to make side-by-side comparison easier ^[8]. The service is also exposed through an MCP server whose read_document tool lets compatible coding agents submit long documents through the same processing layer ^[9].

The people involved suggest the choice was deliberate rather than opportunistic. Pillai holds a robotics PhD from MIT and led machine-learning work at Toyota Research Institute, according to VLM Run's team page ^[10]; founding researcher Dinesh Narapureddy completed a robotics PhD at Carnegie Mellon in 2023 on self-supervising occlusions for computer vision, and the early technical group includes former AWS staff and computer-vision researchers from North Carolina State University ^[11]. That is a group that has watched perception systems fail outside demos.

The cost claim is where the release thins out. VLM 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 ^[12], which works out to under $0.0006 per page ^[13]. The launch article does not include the document set, token assumptions, latency measurements or accuracy scores behind those numbers, and no Gateway-specific pricing or methodology appears in the supplied materials ^[14]. The published pricing page offers a $10 signup balance and a $799 monthly Pro plan carrying $1,000 in included usage, with private deployments, custom rate limits and compliance features for enterprise ^[15] - included usage priced $201 above the fee itself ^[16].

Accuracy decides whether the savings are real. A parser that loses a signature, misreads an invoice total or scrambles a dense table pushes cost into the rest of the workflow ^[17]. VLM Run's own position, that the strongest model varies by language, layout density, scan quality and domain ^[18], is a good argument for routing and a poor substitute for a benchmark.

Watch whether the switching guarantee holds: whether the same document through two models yields differences a downstream system can tolerate, and whether VLM Run publishes the document set and accuracy scores behind the under-$60 figure ^[14][12]. Watch how fast new open-weight releases appear in the catalog, since a neutral routing layer that lags the field is just another integration to rip out ^[7].