01
1. Start with the live model catalog
Choose a model by task and modality, not name recognition. Confirm whether the current catalog entry supports OCR, document Markdown, image chat, video or another visual capability before writing integration code.
Also record the current model ID and displayed pricing. The Gateway is in alpha and the catalog can change, so stale examples should not become hard-coded assumptions.
02
2. Test with the VLM Run CLI
The launch documentation shows uvx vlmrun gw models to list models and uvx vlmrun gw chat -m <model-id> <file> as the basic pattern for document, image and video testing.
Use CLI tests on representative files before application integration. OCR quality can vary on scans, forms, tables, tiny text and multilingual material even when a model performs well on cleaner inputs.
03
3. Connect through the OpenAI-compatible API
VLM Run's current example points the OpenAI client at https://gateway.vlm.run/v1/openai/. Document examples use chat completions with a document_url content item and a Markdown method for compatible models.
Treat request fields as model-specific. Do not assume every catalog model accepts document_url, video, the same context limits or the same output method.
04
4. Evaluate OCR on your own documents
- clean digital PDFs
- scanned pages
- tables and multi-column layouts
- forms and invoices
- small or rotated text
- multilingual documents
Measure extraction accuracy before optimizing only for token or page price. A cheaper model that creates more cleanup work can be more expensive operationally.
05
5. Use Markdown extraction carefully in RAG
Document-to-Markdown can be a useful preprocessing step for retrieval systems because it can retain headings and structural cues better than plain OCR text, depending on the model and document.
Validate page order, table fidelity and failure handling before treating extracted Markdown as authoritative. Keep source-page metadata so downstream answers can cite the original document.
06
6. Check video and MCP support separately
Do not infer video support from image support. Check the live capability matrix and test duration limits, sampling behavior, latency and upload constraints on the exact model you plan to use.
VLM Run also exposes the catalog through MCP for agents. Use the current MCP documentation for setup because the public alpha can evolve faster than copied configuration snippets.
07
7. Confirm the current billing state
The September 2026 Gateway launch page says anonymous access remains available with IP-based rate limits and the Gateway is free during alpha, while published prices are already live as production cost signals.
VLM Run's broader pricing page lists Starter at $0 per month with pay-as-you-go usage and a $10 signup balance. Before scaling, confirm whether your workflow is anonymous alpha traffic, authenticated account usage or a paid tier, and verify the current model rate.
08
8. Production checklist
- pin the intended model ID
- test representative media
- confirm current billing and rate limits
- log model and request metadata
- handle unsupported modality errors
- verify retention and deployment terms
- keep a fallback if a model changes or disappears
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.