You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Responsible Use Agreement
This model has had safety refusals removed. That makes it useful for red-teaming, security research, evaluation, and unfiltered assistant tasks — and also removes guardrails a user must therefore supply themselves.
Prohibited uses (you must agree before access is granted):
- Anything involving the sexual exploitation or endangerment of minors.
- You must be of age 18 years or older to use and download this model.
- You agree any information generated that can cause harm in terms of generating recipe, knowledge to make any materials/substances is your own input and responsibility. You will be accountable for any harm/damage caused by your action/input.
- Content promoting self-harm or suicide.
- Generation of material that is illegal in your jurisdiction, or that targets real individuals for harassment, doxxing, or fraud.
- Any use prohibited by the upstream DeepSeek license.
You are responsible for adding appropriate safety filtering, human review, and access controls for your deployment. The weights are provided as-is, with no warranty. The license is inherited from the upstream DeepSeek base model — review and comply with it before use or redistribution.
Log in or Sign Up to review the conditions and access this model content.
keys-Mac DeepSeek-V4-Flash-0731 MXFP4-MLX Abliterated
Drop-in uncensored / abliterated MXFP4-MLX weights for official DeepSeek-V4-Flash-0731, with DSpark MTP heads kept stock, for a single Mac Studio (M3 Ultra) via oMLX.
Recipe + first-prompt Hermes fix + 1M-context knobs:
drowzeys/keys-Mac-DeepSeek-V4-Flash-DSpark-0731-MXFP4-MLX-Abliterated-49tps
| Base | Vontra/DeepSeek-V4-Flash-0731-MXFP4-MLX (bit-exact 0731, MTP kept) |
| Upstream | deepseek-ai/DeepSeek-V4-Flash-0731 |
| Refusal suite | 32/32 BYPASS (twice) · 0 refuse · 0 garble · 0 empty |
| Ablit | L10–42 · λ=3.5 · k=1 · MTP stock · 33× attn.wo_b · mean Δrel 0.0557 |
| Format | MXFP4 experts + MXFP8 attention (MLX packing). Edited wo_b keeps original e8m0 scales. |
| Runtime | oMLX 0.5.7+ · mtp_enabled · enable_thinking: false · 256 GB unified |
| Context | native 1,048,576 advertised; 41,646-token needle retrieved exactly |
| Speed (stock recipe, same runtime) | count / list / read 44–49 tok/s with DSpark on |
Sister DGX/FP8 pack (same direction): drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-32-32.
⚠️ Responsible Use & gated access (required)
WARNING: This model has had safety refusals removed. That makes it useful for red-teaming, security research, evaluation, and unfiltered assistant tasks — and also removes guardrails you must supply yourself.
Access is gated. By requesting Hugging Face access, downloading, or using these weights, you agree to the terms below (same gate family as the other Keys DSV4F abliterated releases).
See RESPONSIBLE_USE.md for the full agreement text.
Prohibited uses
- Anything involving the sexual exploitation or endangerment of minors.
- You must be of age 18 years or older to use and download this model.
- You agree any information generated that can cause harm in terms of generating recipe, knowledge to make any materials/substances is your own input and responsibility. You will be accountable for any harm/damage caused by your action/input.
- Content promoting self-harm or suicide.
- Generation of material that is illegal in your jurisdiction, or that targets real individuals for harassment, doxxing, or fraud.
- Any use prohibited by the upstream DeepSeek license.
Abliteration
| Method | MXFP8 attn.wo_b dequant → rank-1 project → keep original e8m0 scales, re-encode e4m3 |
| Layers | 10–42 (L0–9 + MTP stock) |
| λ | 3.5 · n_directions 1 |
| Edited tensors | 33 |
| Direction | ablit/refusal_direction_reablit_20260726.npz |
| Meta | ABLIT_META.json |
| Results | results/refusal32-summary.json |
mx.quantize is not used for the edit — it re-derives per-32 scales and adds ~5% Frobenius noise on an untouched tensor.
Download + serve (oMLX)
There is no Docker/GHCR image (Apple Metal). Prebuilts: Homebrew oMLX ≥ 0.5.7 + this weight pack. Full one-shot: AGENTS.md.
# after gated access is approved
brew install python@3.11 huggingface-cli
brew tap jundot/omlx https://github.com/jundot/omlx && brew install omlx
git clone https://github.com/drowzeys/keys-Mac-DeepSeek-V4-Flash-DSpark-0731-MXFP4-MLX-Abliterated-49tps
cd keys-Mac-DeepSeek-V4-Flash-DSpark-0731-MXFP4-MLX-Abliterated-49tps
bash scripts/setup-mac.sh
# pulls this HF repo → ~/models/dsv4f-mxfp4-ablit and wires ~/.omlx
omlx serve --model-dir ~/.omlx/models --host 0.0.0.0 --port 11500
Requests: chat_template_kwargs.enable_thinking = false.
Hermes first-prompt stall (20k system prefix looks “stuck”): disable auxiliary.title_generation, set streaming.enabled: true, enable the 32 GB hot cache, then /new. Details in the GitHub recipe.
Credits
- DeepSeek-AI — DeepSeek-V4-Flash-0731
- Vontra — bit-exact MXFP4-MLX conversion that keeps MTP
- oMLX / jundot — runtime, DSpark MTP, DSA Metal kernels
- Keys (drowzeys) — Mida
wo_babliteration, Mac packing, recipe
License
Inherited from the upstream DeepSeek base model. Review and comply before use or redistribution.
- Downloads last month
- 68
4-bit
Model tree for drowzeys/keys-Mac-DeepSeek-V4-Flash-0731-MXFP4-MLX-Abliterated
Base model
deepseek-ai/DeepSeek-V4-Flash-0731