Baidu Releases Qianfan-VL Vision-Language Models as Open Source for Enterprise Use

Details

• Baidu AI Cloud has open-sourced the Qianfan-VL series, a family of multimodal vision-language models, making both weights and code publicly available.
• The models come in three parameter sizes, each tailored to balance latency, accuracy, and compute requirements for different deployment needs.
• According to Baidu’s technical blog, Qianfan-VL achieves leading benchmark results in optical character recognition and mathematical reasoning tasks.
• APIs are designed for enterprise use-cases, including document parsing, interpreting charts, and seamless integration within the broader Qianfan AI platform.
• Checkpoints can be freely downloaded from GitHub and Hugging Face, under a licence allowing commercial deployment.
• This open release builds on Baidu’s recent initiatives with ERNIE Speed and ERNIE Lite, continuing a pattern of open-sourcing major AI models.
• The models are optimized for mainstream hardware, running efficiently on NVIDIA A100 and H100 GPUs, as well as on China-developed Ascend accelerators.

Impact

Baidu’s open-source move intensifies competition with Meta’s LLaVA-Next and Alibaba’s Qwen-VL, as companies race to boost homegrown multimodal AI stacks. The release could accelerate AI adoption in finance and government sectors by offering cost-effective, OCR-focused solutions. By making these models available globally and aligning with China’s push for ‘secure and controllable AI,’ Baidu aims to attract diverse developers and shift industry focus toward application-driven innovation.