AutomatosX/AX-Unlimited-OCR-2-CUDA-AXQ-NVFP4-W4A16

AX Unlimited-OCR-2 CUDA AXQ NVFP4 W4A16

Development storage export. No compatible inference runtime has been verified.

This is an AXQuant RTN NVFP4 W4A16 export of baidu/Unlimited-OCR-2, from original BF16 weights at revision 9cf077c4b206e0e2a31879df937b0b0bcba3661c. The source weight SHA-256 digests were checked against the pinned Hub LFS objects. This pack is provided for format development and research.

Storage

• 2,196 language attention, MLP and expert matrices use NVFP4 E2M1 weights, E4M3FN block scales (group size 16) and FP32 global scales.
• 623 other tensors retain their exact source values and dtypes, including all 108 SR-SWA DASH-router/indexer tensors, vision encoders, projector, normalization, embedding and output-head state.
• Activations are declared BF16 (W4A16); activation quantization is disabled.
• The explicit CPU reference encoder produced the checkpoint: backend=numpy-reference, status=development, runtime_verified=false, quality_certified=false.
• axquant_cuda_plan.json binds the source files and per-tensor allocations. axquant_cuda_manifest.json binds the plan and emitted payloads.

Runtime limitations

Unlimited-OCR-2 adds SR-SWA sparse routing, page-aware rotary positions and multi-page state. Its architecture differs from Unlimited-OCR v1 despite sharing model_type=unlimited-ocr.

No CUDA, vLLM, SGLang, Transformers or AX Engine load/generation test has been run for this exact pack. The upstream BF16 inference commands cannot be assumed to consume these packed tensor names. A runtime needs both the second-generation graph and the NVFP4 storage integration. There is currently no verified inference command for this repository.

Storage layout, source hashes, preserved tensor equality and complete output index coverage were checked. These checks do not establish OCR accuracy, long-document functionality, throughput, memory bounds or certification. No MTP or acceleration claim is made. MLX MXFP4/MXFP8 versions are not included.

License and upstream notices

The upstream model card declares MIT. The included LICENSE preserves Baidu's MIT notice from its official Unlimited-OCR source repository (Git blob 890d455ae73d1d930eee703ce1e5478783ec9154). Original model code files and their notices are preserved; they are not adapted to consume NVFP4 tensors. The pinned source card is included as UPSTREAM_README.md.