Hosted vision-language models

Vision-language models process both images and text, enabling tasks such as extracting structured data from scanned documents, answering questions about photographs, and generating captions for accessibility. For example, deepseek-ai/DeepSeek-OCR-2 is specialised for optical character recognition, while series like Qwen/Qwen2.5-VL-7B-Instruct and Qwen/Qwen2-VL-2B-Instruct support visual question answering and image-to-text generation.

• Document digitisation and invoice parsing
• Automated content moderation on visual platforms
• Visual search and retrieval-augmented generation (RAG) pipelines

In production, these models are often integrated into RAG workflows or multimodal chatbots. Choosing among the 32 models listed here involves balancing latency, accuracy, and cost: larger architectures such as Qwen/Qwen3.6-35B-A3B-FP8 yield higher quality on complex reasoning but require more compute, while quantised or smaller models like cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit or Qwen/Qwen3-VL-4B-Instruct serve well at lower throughputs. For most call volumes, calling a hosted API eliminates infrastructure overhead and enables elastic scaling — benefits gigarouter provides through its benchmarked, OpenAI-compatible endpoints. (Currently 0 models are live; the remainder are being onboarded.)