Skip to main content
The VLM Run Gateway is a fully OpenAI chat completions-compatible API for visual intelligence. If you’re building document extraction or visual understanding, Gateway exposes OCR, VQA, and detection behind a single interface you already know: OpenAI’s chat.completions. Point base_url at https://gateway.vlm.run/v1/openai, and every model on the catalog speaks the same request shape your existing SDK calls already use.

​
Why the Gateway

General-purpose model routers are built for text LLMs, so they usually cover only a small slice of visual workloads. The VLM Run Gateway is built around OCR and VQA instead:
  • One API, multiple supported models. Use qwen/qwen3.5-0.8b for visual question answering over images and video, paddleocr/pp-ocrv6 for OCR and text detection, zai-org/glm-ocr / rednote-hilab/dots.mocr for document OCR and layout parsing, and paddlepaddle/pp-doclayoutv3 for layout detection with reading order, all through the same chat.completions shape.
  • Embeddings and transcription too. qwen/qwen3-vl-embedding-2b embeds text and images into the same vector space for visual search, and nvidia/parakeet-tdt-0.6b-v3 handles audio transcription, including subtitle output (srt/vtt). Both are available through their own OpenAI-compatible endpoints under the same base_url and API key.
  • Document routing built in. Multi-page PDFs are decoded, rasterized, and fanned out per page automatically at the ingress layer, so you do not have to split or stitch documents yourself.
  • Method and DPI routing. method selects the operation a model runs (ocr, detect, markdown, …); document_dpi tunes how much detail reaches the OCR model per page.
  • Operational signals included. Use the live catalog at GET /v1/openai/models, per-request usage.cost for metering, and x-request-id for tracing on every response.
  • Drop-in for agents. The MCP server exposes read_document, read_image, read_audio, and read_video tools over Streamable HTTP, so any MCP-aware framework (Pydantic AI, LangChain, Mastra, OpenAI Agents SDK, Claude Code) can read files through the same pipeline, auth, and billing.
Treat model, method, and DPI as one tunable unit: start with the cheapest combination that clears your accuracy bar, and only pay for more where it moves the numbers.

​
Next steps

Quickstart

Prerequisites, plus first VQA and document OCR requests.

Flexible Document OCR

Request knobs, page blocks, and streaming for PDFs.

Models

Catalog, capabilities, and model selection.

Methods

Method and method_params reference.

MCP Server

Connect any MCP-aware agent (Pydantic AI, LangChain, Mastra, Claude Code).