Released 2026-06-22 · document vision
Deployment evidence
Baidu · 3B total · 3B dense active · 32K tokens
Best-fit scenarios
Compact document-vision model for one-shot long-document OCR, multi-page PDF parsing and structured layout extraction.
Not recommended
- Sending unbounded PDFs without page, pixel and token admission limits
- Assuming OCR benchmark scores cover every language, handwriting or layout
Long-document OCR
Vendor-statedBaidu positions the model for one-shot long-horizon OCR rather than page-fragment-only extraction.
Capability sourceMulti-page PDF parsing
Vendor-statedThe official processor and examples cover document images and PDF-oriented workflows with a 32K generation configuration.
Capability sourceStructured layout extraction API
ChinaAPI inferenceThe official vLLM and SGLang OpenAI-compatible paths make queued document extraction plausible after processor and post-processing validation.
Official launch / minimum
Evidence COfficial single gpu path no vram floor
The official Transformers example loads the model on one CUDA GPU, but the vendor does not publish an absolute minimum VRAM configuration. Primary recipe
200-person team
Evidence EDocument queue benchmark required
Size by pages, pixels, document length and peak jobs rather than seats; start with one measured queue worker.
Commercial API
Evidence COfficial framework paths
Official vLLM and SGLang serving paths exist; production needs replicated workers, document limits, processor pinning and post-processing observability. Primary recipe
Commercial-use check
MIT
MaaS / hosted service: No model-specific MaaS restriction identified in the MIT license.
Attribution: Retain the copyright and permission notice in copies or substantial portions of the weights.
Known limitations and open questions
- No ChinaAPI hardware or OCR-quality reproduction
- PDFs still require deterministic image conversion and document post-processing
- The 32K setting is taken from the official inference configuration, not a full-context concurrency result