Qianfan-VL-3B

Qianfan-VL-3B brings the Qianfan multimodal recipe to a smaller footprint suited to edge and high-throughput scenarios. It accepts images alongside prompts—scanned pages, receipts, charts, screenshots, or product photos—and produces concise, grounded text that follows instructions reliably. While it trades some peak accuracy for responsiveness, it maintains layout-aware reading, handles small text competently, and keeps references straight across multiple images or pages. The model supports streaming, long contexts, and function calling, enabling agents to crop regions, retrieve context, or format results as JSON without complex glue code. Deployed on Baidu’s Qianfan stack, it slots into production with the same APIs and guardrails as larger tiers. Teams adopt the 3B variant for lightweight document workflows, screenshot and UI helpers, multimodal search, and real-time assistants where low latency and cost matter most.