Downloads · 30 days
117
3% of all-time downloads
bullerwins/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-32-32-GGUF
keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-32-32-GGUF is a text generation model from bullerwins. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as other.
Drop-in uncensored / abliterated weights for official DeepSeek-V4-Flash-0731 (GA), with DSpark MTP modules kept stock, for dual DGX Spark / Anemll + MiaAI-Lab serve.
Downloads · 30 days
117
3% of all-time downloads
All-time downloads
3.9K
Public
Repo size
173 GB
Likes
4
Public
Click a slice to open those files.
.gguf173 GB · 100%
From the Hugging Face model README
Drop-in uncensored / abliterated weights for official DeepSeek-V4-Flash-0731 (GA), with DSpark MTP modules kept stock, for dual DGX Spark / Anemll + MiaAI-Lab serve.
| Base | deepseek-ai/DeepSeek-V4-Flash-0731 @ 9e165c30 |
| Refusal suite | 32/32 BYPASS (QuantTrio-style hard suite) · 0 refuse · 0 garble · 0 errors |
| Ablit | L10–42 · λ=3.5 · k=1 · MTP stock (33× attn.wo_b) · mean Δrel ≈ 0.0556 |
| Scales | float8_e8m0fnu preserved (e8m0) |
| Runtime | Anemll ghcr.io/anemll/dspark-vllm-gx10:0.1.1 + MiaAI 2× TP=2 · MTP=5 · nvfp4_ds_mla · util ≤ 0.85 |
Full credit: Anemll/dspark-vllm-gx10 · MiaAI-Lab/DeepSeek-v4-Flash-DSpark-2x-DGX-Spark · DeepSeek-AI
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.
| Field | Description |
|---|---|
| Username | Your name or handle (form may default to your HF username) |
| Contact email (form may default to your HF account email) | |
| Reason for intended use | e.g. red-teaming, security research, evaluation, local assistant |
Plus all agreement checkboxes under Prohibited uses.
| Method | Layer-range FP8 attn.wo_b SRA / projection |
| Layers | 10–42 (L0–9 + MTP stock) |
| λ | 3.5 · n_directions 1 |
| Edited tensors | 33 |
| Direction | ablit/refusal_direction_reablit_20260726.pt |
| Meta | ABLIT_META.json |
| Path | Purpose |
|---|---|
model-*-of-00048.safetensors + index | Full FP8 checkpoint (incl. embedded DSpark MTP stock) |
encoding/encoding_dsv4.py | 0731 tokenizer encoding helper for Anemll/Mia serve |
ABLIT_META.json | Edit stats / recipe fingerprint |
ablit/refusal_direction_reablit_20260726.pt | Direction used for this reablit |
ablit/refusal_direction_r1.pt | Related SRA direction artifact |
results/refusal32-clean-*.json | Live 32/32 suite summary |
# after access is approved
hf download drowzeys/keys-DeepSeekV4-Flash-GA-0731-Dspark-Abliterated-32-32 \
--local-dir ~/models/DeepSeek-V4-Flash-0731-ablit-100
docker pull ghcr.io/anemll/dspark-vllm-gx10:0.1.1
# Point MiaAI DeepSeek-v4-Flash-DSpark-2x-DGX-Spark-0731 recipe at the local dir
# DSPARK_MODEL=... SERVED_MODEL_NAME=deepseek-v4-flash-0731-ablit-100
# MTP_NUM_TOKENS=5 GPU_MEMORY_UTILIZATION≤0.85 MAX_MODEL_LEN=1048576
Refusal: 32/32 BYPASS.
Decode (short monologue, accept ~0.3): C1 ~31 · C4 ~71 · C6 ~87 tok/s aggregate.
Code-like prompts (accept ~0.7): C1 ~46–65 tok/s.
Stage-C knobs (local argmax / Markov / etc.) are not exposed on Anemll 0.1.1; historical champion peaks used Stage-C runtime.