keys-Mac-oMLX-0.6.3.2RC-DeepSeekV4F-0731-MXFP4-Abliterated-Dual-ANE-CPU

1.0 Beta · Mac Studio M3 Ultra · oMLX 0.6.3rc2 (pack name 0.6.3.2RC) · DeepSeek-V4-Flash 0731 MXFP4 abliterated · Dual-ANE + CPU prefill · DSpark MTP · 1M context

Previous 49 tok/s / oMLX 0.5.7 pack: previous-version/

These weights have safety refusals removed. Research / red-team only — you supply the guardrails.

This recipe stands on other people's work. Please star them.

• jundot/omlx — oMLX, MTP/DSpark on Mac, Dual-ANE prefill
• DeepSeek-AI — V4-Flash 0731 + DSpark
• Vontra MXFP4-MLX — 4-bit Mac quant
• ml-explore/MLX
• Spark cousins: Anemll, MiaAI-Lab, Tony / tonyd2wild

Full list: CREDITS.md. Please donate / support: GoFundMe.

Keep 4-bit MXFP4. Dual-ANE is prefill-only on DSV4 (decode is noise vs ANE-off).

First request after load can be slower (~356). Prefix-cache second runs (~1114 tok/s on 5.6k) are not Dual-ANE wins.

count 48.9 / list 44.8 / read 49.7 tok/s on 4-bit MXFP4. Different protocol — do not mix with the table above.

git clone https://github.com/drowzeys/keys-Mac-oMLX-0.6.3.2RC-DeepSeekV4F-0731-MXFP4-Abliterated-Dual-ANE-CPU.git
cd keys-Mac-oMLX-0.6.3.2RC-DeepSeekV4F-0731-MXFP4-Abliterated-Dual-ANE-CPU
# Python 3.11, oMLX 0.6.3rc2, gated HF pack
bash scripts/setup-mac.sh
omlx serve --model-dir ~/.omlx/models --host 0.0.0.0 --port 11500

Agent recipe: AGENTS.md

GHCR carrier (recipe + benches, not a Metal runtime):

docker run --rm -v "$PWD":/out ghcr.io/drowzeys/keys-mac-dsv4f-mxfp4-dualane:1.0-beta \
  cp -r /omlx/. /out/

Aday777 NVFP4 Dual-ANE is a no-op on this path. Stay on MXFP4.

Thanks to everyone who contributed — full credits in CREDITS.md. Pack: drowzeys / keys.