VLM Run Gateway Launches Unified API for Open-Weight OCR and Vision Models
What shipped
VLM Run Gateway launched September 4, 2026 (announced on the Hugging Face blog): one API for running open-weight VLMs, OCR VLMs, and ViT-based vision models, built from the team's own production pain serving vision for agents and evals.
What changed
- Pinned servicing — a model ID means one exact quantization and serving config, killing the silent quality drift from providers swapping quants, which hits OCR and small-text accuracy hardest.
- Video done right — video-native VLMs with controllable FPS sampling; most routers don't support video inputs or time resolution at all.
- Document inference built in — PDF rasterization, parallel page workers, and retries handled for you.
- Swap-by-name catalog — GLM-OCR, dots.mocr, PaddleOCR VL 1.6, Qwen3.8-27B, Gemma4-26B-A4B, and more behind one interface.
Why a builder cares
Vision quality degrades silently through generic routers — same model ID, different quant, dropped video FPS, no page-level retries. VLM Run Gateway makes open-weight vision behave like one dependable API, which is exactly what OCR-heavy agents need.