AutomatosX/AX-Unlimited-OCR-2-AXQ-MXFP8

AX Unlimited-OCR-2 AXQ MXFP8 reference storage

Development reference storage. Full model inference is not supported.

Original source: baidu/Unlimited-OCR-2 at revision 9cf077c4b206e0e2a31879df937b0b0bcba3661c. Original BF16 weights were checked against pinned Hub LFS SHA-256 objects. This pack was exported using the AXQuant development workspace for research and format development.

Storage

2,196 language attention/MLP/expert matrices use MLX-native MXFP8 encoding with group size 32. Packed arrays keep the original .weight name as U32; .scales stores one U8 E8M0 exponent per group. No affine biases are present. Individual experts retain their original names and layout.

All other 623 tensors retain exact source values and dtypes, including all 108 SR-SWA DASH-router/indexer tensors, vision encoders, projector, norms, embedding and output head. The complete source config is preserved with added packing metadata and an explicit axquant_storage_only marker.

Weight files occupy 4.167 GB, or 9.969099 measured storage BPW including protected weights, scales and Safetensors headers. Nominal format bits are not whole-pack BPW.

axquant_plan.json, axquant_source_binding.json and SOURCE_SHA256SUMS.txt record the plan, structural source identity and source member hashes. axquant_manifest.json binds the output files and static compatibility report. The manual plan is explicitly unmeasured; quantization is not calibration or a quality result.

Validation and limits

• Checked all 5,015 indexed output tensors and their storage shapes/dtypes.
• Checked exact equality for every preserved tensor and all source config fields.
• Independently decoded 24 matrix samples and compared them with public MLX dequantization and quantized matrix multiplication. This verifies matrix storage/operations only.
• oMLX and MTPLX static profiles report unsupported. Primary runtime metadata has standard_inference=false.

Unlimited-OCR-2 requires an SR-SWA graph, page-aware rotary positions and multi-page KV lifecycle that are not implemented in the current MLX runtime path. No MLX-LM, MLX-VLM, AX Engine, oMLX or MTPLX full-model load/generation was run. No OCR quality, long-document behavior, memory bound, speed or MTP claim is made. No quality certification is attached.

There is no supported model inference command for this repository. Included upstream Python files are unchanged archival source assets; their BF16 loader cannot be assumed to consume this packed layout. Future runtime integration needs its own layout validation and exact-pack tests.

License

The upstream model card declares MIT. LICENSE preserves Baidu's official MIT notice (source repository Git blob 890d455ae73d1d930eee703ce1e5478783ec9154). Original code notices are retained. The pinned source model card is included as UPSTREAM_README.md.