AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit

Development / experimental AXQuant 2-bit pack of deepseek-ai/DeepSeek-V4-Flash-0731 @ 7872f01b1d1fe23eabc4c98b48bffcef5a386062.

Converted on df-macstudio-m2 (Apple M2 Ultra, 192 GB) from the native FP8 0731 source (quant_method=fp8). Product class 2bit-experimental.

This is not the older DeepSeek-V4-Flash Hub pack. Do not treat AX-DeepSeek-V4-Flash-MLX-AXQ-2bit certificates as evidence for this 0731 revision.

Measured precision

Property	Value
Target class	2bit-experimental
Measured main BPW	3.1328993873020314
Measured total BPW	3.2142055528774454
Weight bytes	122,212,298,775
Source	deepseek-ai/DeepSeek-V4-Flash-0731@7872f01b1d1fe23eabc4c98b48bffcef5a386062
Convert host	df-macstudio-m2
AXQuant	1.8.1
mlx / mlx-lm	0.32.0 / 0.31.3 (vendored deepseek_v4 + FP8 load hook)

Claims

Claim	Status
Converted on Studio from the pinned 0731 revision	Yes
mlx-lm load + generate smoke	Passed on df-macstudio-m2
Official DSV4 chat_template.jinja	In pack
Checkpoint Tier 1 (generation viability suite)	Not certified on this record; Studio 15+15 factory QA with chat is 0.633 combined
AX Engine native manifest	Not generated — generate-manifest --validate rejected split switch_mlp.gate_proj / up_proj ([256, 2048, 256] vs fused [256, 4096, 256])
MTP acceleration	Not certified

Requires AX_ENGINE_2BIT_EXPERIMENTAL=1 if served with AX Engine after a future manifest fix. mlx-lm generate does not need that env.

Attribution

Base weights © DeepSeek. Quantization by AXQuant (development).