VLM Run opens a unified gateway for open-weight vision models

VLM Run has launched Gateway, a unified API for open-weight OCR, vision-language and ViT-based models. The alpha service is currently free and requires no sign-up. The announcement was published by the Hugging Face Blog on September 4, 2026.

For AI developers, the appeal is less about adding another model endpoint and more about reducing the operational differences between providers. VLM Run says vision workloads are particularly sensitive to quantization, serving parameters, maximum resolution, context limits and GPU architecture—variables that can change OCR, small-text and spatial accuracy even when the model name stays the same.

Video is another focus. In the company’s testing of popular model routers, more than 80% of providers did not support video inputs, while fewer offered control over frames-per-second sampling. Gateway aims to expose models with those capabilities through a consistent interface.

The service also handles document-processing work that usually falls on application teams, including PDF rasterization, parallel page execution, retries for failed pages and rate-limit management. Developers can switch models by changing a model name, with examples including GLM-OCR, DeepSeek-OCR-2, dots.mocr, PaddleOCR VL 1.6, Qwen3.8-27B and Gemma4-26B-A4B.

A command-line client is available through pip install vlmrun or uvx. Users can list available models, analyze PDFs and images, or send video prompts—for example, uvx vlmrun gw chat <vid>.mp4 -m qwen/qwen3.5-0.8b -p "describe the video".

For teams building multimodal agents and evaluations, the gateway could make model comparisons faster while shifting serving and pipeline maintenance to the platform. It remains an early alpha, so developers should still validate visual quality for their specific workloads.

Source: Hugging Face Blog