Avdpro/DeepSeek-V4-Flash-2bit-DQ-SSD

AI2Apps SSD-ready checkpoint

This is a byte-preserving storage-layout conversion of mlx-community/DeepSeek-V4-Flash-2bit-DQ at immutable revision 722bf559b7de93575b2320973cf2002e05bfe6c9. Original tensor values and quantization are retained. Original routed experts are externalized into the experts/ directory rather than duplicated in the backbone safetensors.

Requires an AI2Apps Runtime with explicit deepseek-v4-expert-major-v1 support. This candidate is not a drop-in checkpoint for unmodified Transformers, mlx-lm or mlx-vlm. Do not use the backbone safetensors alone. The corresponding Runtime and model Package have not yet completed release acceptance. Full/Cached engine compatibility is recorded separately in the Runtime release receipt; the presence of a reversible tensor map alone is not an end-to-end engine guarantee.

• ssd-checkpoint.json: format, provenance and file digests.
• external-tensors.json: original tensor names and external byte locations.
• source-tensor-sha256.json: original tensor payload digests verified during export.
• model.safetensors.index.json: ordinary/vision/other retained tensor index.
• experts/: complete routed expert payloads, with no re-quantization.

See LICENSE and README.upstream.md for upstream terms and attribution. This storage format changes installation space and data access; it is not a new model training or a claim of improved model accuracy.