Model request: MinerU2.5-Pro-1.2B (document parsing VLM)

Model

Name: MinerU2.5-Pro-2604-1.2B
Architecture: Qwen2VL (same as qwen2.5vl-it:3b which already works)
Size: 1.2B parameters
Use case: Document parsing — converts PDFs/images to structured Markdown/JSON with layout detection, table extraction, formula recognition, and OCR

Why this model

• MinerU is a widely-used document extraction engine (14k+ GitHub stars)
• MinerU2.5-Pro scores 95.69 on OmniDocBench v1.6 (SOTA for document parsing)
• At 1.2B it's well within XDNA2 NPU capacity — qwen2.5vl-it:3b already runs successfully on Strix Halo
• MinerU supports an http-client backend that points at any OpenAI-compatible API server — flm serve would slot in directly
• Running document parsing on the NPU at <2W would be a compelling use case for edge AI

Hardware tested

• Strix Halo (gfx1151), 8 NPU columns, firmware 1.1.2.65, amdxdna 0.6
• qwen2.5vl-it:3b loads and runs correctly on this hardware

Notes

Since MinerU2.5 uses the same Qwen2VL architecture as the already-supported 3B model, the existing libqwen2vl_npu.so library and xclbin kernels may work with just a q4nx weight conversion — assuming tensor dimensions are compatible.

Start by examining the existing qwen2.5vl-it:3b support, especially libqwen2vl_npu.so and its xclbin kernels. Check whether MinerU2.5-Pro-2604-1.2B can use the existing Qwen2VL path after q4nx conversion and run through flm serve; done means the model loads and performs document-parsing inference on the tested hardware.