VLM Run opens a vision-model gateway for OCR, documents and video

Why it matters

VLM Run is betting that visual AI's durable business sits in orchestration: stable APIs, document pipelines and video controls that survive rapid model turnover.

Sudeep Pillai opened VLM Run Gateway to the public on September 1st, giving developers one OpenAI-compatible interface for open-weight OCR, vision-language and specialized computer-vision models. VLM Run followed with a September 4th community article on Hugging Face that laid out the founder's case: serving a vision model is easy to demo and considerably harder to trust once PDFs, video and production accuracy enter the request.

Pillai came to that problem after building perception systems rather than chat applications. He earned a Ph.D. at MIT with a thesis on self-supervised perception and SLAM in mobile robots, then worked as an ML lead at Toyota Research Institute, where he shipped perception systems for Toyota's autonomous-vehicle research.

That background explains the part of Gateway that VLM Run is choosing to own. The product does not simply route requests to a list of models. VLM Run handles PDF rasterization, parallel page processing, retries, rate limits and normalized outputs, including Markdown, bounding boxes and structured JSON. For video, VLM Run exposes controls such as frame sampling that determine how much temporal information reaches the model.

The public alpha is free and requires no sign-up. Developers can point an OpenAI client at the Gateway endpoint, then switch among models by changing the model parameter rather than rebuilding an inference stack for each checkpoint.

Pillai is selling reversibility

VLM Run's sharper bet is that production teams will need to change visual models frequently. OCR accuracy varies with page layout, language, scan quality and handwriting. Video models vary in their support for native video inputs and frame-rate controls. A general vision-language model that performs well on open-ended image reasoning can remain an expensive or inaccurate choice for reading a dense table or low-resolution fax.

Gateway currently presents models for OCR, document parsing, image and video chat, detection, captioning, embeddings and transcription. The model catalog includes GLM-OCR, dots.mocr, PaddleOCR-VL, Qwen and Gemma variants, among others. The interface also extends to an MCP server, allowing compatible agent frameworks to invoke document, video and audio tools through the same access and billing layer.

Pillai and VLM Run founding researcher Dinesh Narapureddy wrote that Gateway grew out of infrastructure built for VLM Run's Orion visual agent and internal evaluations. Orion can call multiple underlying models during one request because localization, segmentation, OCR and video understanding do not share a single consistent winner. Opening those models directly turns internal infrastructure into a separate developer product.

That move also changes VLM Run's position in the stack. Orion packages visual reasoning into an agent that can inspect and act on images, video and documents. Gateway moves lower, offering the serving and orchestration layer to developers who want to build their own agents or extraction systems. Customers can use VLM Run's agent or assemble competing applications on VLM Run's infrastructure.

RuntimeWire reported in August that VLM Run had placed six open OCR models behind one compatible endpoint. The September release widens that idea beyond OCR and makes the strategic wager clearer: individual models will keep changing, while the surrounding document and video plumbing can become the durable product.

Model IDs hide production choices

VLM Run says providers frequently serve different quantizations, runtimes and inference parameters under the same model ID. Those implementation choices can affect OCR, small-text recognition and spatial accuracy even when text benchmarks show little movement. Gateway is Pillai's attempt to make the deployed configuration part of the service rather than an invisible decision made by an upstream provider.

VLM Run also says its tests of popular routers found that more than 80% of providers did not accept video inputs for video-native models, with fewer still offering frame-rate control. That result is VLM Run's own testing and lacks an independently published comparison. The underlying product problem is concrete: support for a model name does not guarantee support for every input or capability the weights can handle.

The same distinction applies to VLM Run's performance and cost comparisons. In its September 1st launch post, VLM Run published internal benchmark results arguing that specialized open-weight OCR and video models can undercut commercial parsers and frontier APIs on selected workloads. Those results were produced by VLM Run and should be read as product evidence, not a neutral ranking. Gateway's practical value will depend on reliability, latency and output quality on each customer's data.

A small team moves down the stack

VLM Run dates to 2023 and lists backing from South Park Commons, Gallery and York IE, without publishing a financing amount. South Park Commons identifies Pillai as a Founder Fellow and places VLM Run in San Francisco. VLM Run's current team page also names Narapureddy, a Carnegie Mellon robotics Ph.D. and former AWS AI Labs scientist, alongside researchers and engineers with backgrounds in computer vision and production machine learning.

VLM Run says Orion processes hundreds of thousands of requests spanning millions of tool calls each month across healthcare claims, construction submittals, contracts and faxes. VLM Run has not published independently verified revenue, customer-count or usage data. The broader pricing page shows how the free alpha can feed the commercial business: paid plans add throughput, included usage, zero-data retention, business associate agreements and private-cloud deployment.

Gateway gives Pillai a second route into visual-AI budgets. A developer may arrive to compare OCR models for a PDF workflow, then need the compliance, observability and deployment controls attached to VLM Run's paid platform. Free access reduces the cost of that first experiment while model usage gives VLM Run direct evidence about which open-weight systems developers are willing to run in production.

Pillai's bet is grounded in a familiar infrastructure pattern. Model builders compete to produce the best checkpoint, while application teams still need stable interfaces, measurable costs and predictable output contracts. Gateway asks developers to let VLM Run own that layer for visual workloads. The models can keep moving. Pillai wants the endpoint to stay put.