0
stars
8
commits
1
linked in READMEs
Aug 21, 2026
updated
README
AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit-MTP
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.
Recipe: AXQuant manual
deepseek-v4-experimental-2bit-v0.1.yaml
(uniform 2-bit trunk). MTP sidecar is packaged (mtp.safetensors).
This is not the older
DeepSeek-V4-FlashHub pack. Do not treat AX-DeepSeek-V4-Flash-MLX-AXQ-2bit-MTP certificates as evidence for this 0731 revision.
Recipe (uniform v0.1)
This is the AXQuant assignment that scored best among 2-bit-class converts
on the factory v-extract suite. Later mixed / attention-6 / shared-4-bit
recipes scored worse and are not this pack.
| Tensors | Bits | Method |
|---|---|---|
| Routed experts + MLP | 2 | affine, group 32 |
| Attention | 4 | affine, group 32 |
| Embeddings / routers | 8 | affine, group 32 |
| Norms, LM head, MTP | 16 | bf16 |
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 | 1.9.0 |
Claims
| Claim | Status |
|---|---|
| Converted on Studio from the pinned 0731 revision | Yes |
Official DSV4 chat_template.jinja | In pack |
| Checkpoint Tier 1 (generation viability suite) | Not certified — 7.1.5 native 15+15 combined 0.633; v-extract on AX Engine HEAD 80f2a3e6 combined 0.887 (floor 0.90). Distinct 2-bit recipe converts scored worse. |
| AX Engine 7.1.5 native load | Passed on df-macstudio-m2 (Hub commit cb1a34b4, --stream-experts off, chat smoke Okay.) |
| Decode-128 (informational) | 15.535 tok/s on 7.1.5; not a Tier 1 claim |
MTP assets (mtp.safetensors) | Packaged — Hub name uses -MTP |
| MTP acceleration | Not certified (T1 below 0.90; MTP A/B not run) |
Requires AX_ENGINE_2BIT_EXPERIMENTAL=1 for AX Engine native serve.
Certificate:
deepseek-v4-flash-0731-axq2-tier1.md.
Comparison vs OptiQ 2-bit:
optiq2-vs-axq2-v190.
Runtime and MTP policy
Download the complete snapshot before serving it:
hf download AutomatosX/AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit-MTP \
--local-dir ./AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit-MTP
AX_ENGINE_2BIT_EXPERIMENTAL=1 \
ax-engine serve ./AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit-MTP --port 31418
mtp.safetensors is the native DeepSeek V4 nextn sidecar. It is not a Qwen
qwen3-next-mtp sidecar, so the oMLX/MTPLX Qwen import workflow does not apply. AX Engine 7.1.5
recognizes the sidecar, but keeps this checkpoint on direct fallback because no revision-bound
Tier 2 MTP acceptance, exactness, or speed evidence exists. The internal DeepSeek MTP
certification-candidate switch is for the formal harness, not normal serving. Stock MLX-LM can
run the text backbone but does not activate this sidecar.
Attribution
Base weights © DeepSeek. Quantization by AXQuant (development).
Contributors
8 commits
Similar repos
0
stars
8
commits
1
linked in READMEs
Aug 21, 2026
updated
README
AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit-MTP
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.
Recipe: AXQuant manual
deepseek-v4-experimental-2bit-v0.1.yaml
(uniform 2-bit trunk). MTP sidecar is packaged (mtp.safetensors).
This is not the older
DeepSeek-V4-FlashHub pack. Do not treat AX-DeepSeek-V4-Flash-MLX-AXQ-2bit-MTP certificates as evidence for this 0731 revision.
Recipe (uniform v0.1)
This is the AXQuant assignment that scored best among 2-bit-class converts
on the factory v-extract suite. Later mixed / attention-6 / shared-4-bit
recipes scored worse and are not this pack.
| Tensors | Bits | Method |
|---|---|---|
| Routed experts + MLP | 2 | affine, group 32 |
| Attention | 4 | affine, group 32 |
| Embeddings / routers | 8 | affine, group 32 |
| Norms, LM head, MTP | 16 | bf16 |
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 | 1.9.0 |
Claims
| Claim | Status |
|---|---|
| Converted on Studio from the pinned 0731 revision | Yes |
Official DSV4 chat_template.jinja | In pack |
| Checkpoint Tier 1 (generation viability suite) | Not certified — 7.1.5 native 15+15 combined 0.633; v-extract on AX Engine HEAD 80f2a3e6 combined 0.887 (floor 0.90). Distinct 2-bit recipe converts scored worse. |
| AX Engine 7.1.5 native load | Passed on df-macstudio-m2 (Hub commit cb1a34b4, --stream-experts off, chat smoke Okay.) |
| Decode-128 (informational) | 15.535 tok/s on 7.1.5; not a Tier 1 claim |
MTP assets (mtp.safetensors) | Packaged — Hub name uses -MTP |
| MTP acceleration | Not certified (T1 below 0.90; MTP A/B not run) |
Requires AX_ENGINE_2BIT_EXPERIMENTAL=1 for AX Engine native serve.
Certificate:
deepseek-v4-flash-0731-axq2-tier1.md.
Comparison vs OptiQ 2-bit:
optiq2-vs-axq2-v190.
Runtime and MTP policy
Download the complete snapshot before serving it:
hf download AutomatosX/AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit-MTP \
--local-dir ./AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit-MTP
AX_ENGINE_2BIT_EXPERIMENTAL=1 \
ax-engine serve ./AX-DeepSeek-V4-Flash-0731-MLX-AXQ-2bit-MTP --port 31418
mtp.safetensors is the native DeepSeek V4 nextn sidecar. It is not a Qwen
qwen3-next-mtp sidecar, so the oMLX/MTPLX Qwen import workflow does not apply. AX Engine 7.1.5
recognizes the sidecar, but keeps this checkpoint on direct fallback because no revision-bound
Tier 2 MTP acceptance, exactness, or speed evidence exists. The internal DeepSeek MTP
certification-candidate switch is for the formal harness, not normal serving. Stock MLX-LM can
run the text backbone but does not activate this sidecar.
Attribution
Base weights © DeepSeek. Quantization by AXQuant (development).
Similar repos
Linked in READMEs
Contributors
8 commits