license: mit
base_model: deepseek-ai/DeepSeek-V4-Flash-DSpark
tags:
- deepseek
- deepseek-v4
- dspark
- abliterated
- uncensored
- moe
- nvfp4
- vllm
- dgx-spark
- v1.1-alpha
- mida-brikie
library_name: transformers
pipeline_tag: text-generation
DeepSeek-V4-Flash-DSpark — Abliterated Uncensored v1.1 alpha (Mida / Brikie)
Mida / Brikie / Hermes-friendly abliteration (milder than v1.0).
Not the same weights as v1.0 alpha.
| v1.1 alpha (this repo) | v1.0 alpha | |
|---|---|---|
| HF | this model | Abliterated-Uncensored |
| Recipe | L10–42 wo_b · MTP stock · λ=3.5 |
L10–42 + MTP · λ=3.5 |
| C1 pure | ~50 tok/s | ~57 tok/s |
| Standing concurrency | C=4 @ 1M | often C12 in publish notes |
| Refusal 32/32 | 100% | 100% |
| Agent surface | less skill-catalog spill | more spill under fat Hermes indexes |
GitHub (both releases):
https://github.com/drowzeys/DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored-1M-57toks
→ RELEASES.md
⚠️ 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 as v1.0).
Access request fields
| 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.
Full text: repo RESPONSIBLE_USE.md
Prohibited uses
- Anything involving the sexual exploitation or endangerment of minors.
- You must be 18 years of age or older to use and download this model.
- Any information generated that can cause harm (recipes / knowledge to make materials or substances) is your own input and responsibility; you are accountable for any harm/damage from your action/input.
- Content promoting self-harm or suicide.
- 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 must add safety filtering, human review, and access controls. Weights as-is, no warranty.
Method (v1.1)
- mHC-resistant direct FP8
attn.wo_bprojection (not LoRA). - Stock
wo_bon layers 0–9. - Abliterated
wo_bon layers 10–42 only. - MTP draft heads left stock (
--no-mtp) — protocol / agent friendliness. - SRA rank-1 refusal direction · λ = 3.5.
Base: deepseek-ai/DeepSeek-V4-Flash-DSpark
Measured (2× DGX Spark GB10, TP=2)
| metric | value |
|---|---|
| Context | 1,048,576 |
| KV | nvfp4_ds_mla · pool ~2.89M @ UTIL 0.85 |
| Standing max_num_seqs | 4 |
| C1 server agg (code) | ~50 tok/s |
| C4 server agg | ~113 tok/s |
| Refusal suite | 32/32 bypass · 0 garble |
Serve (C=4 · 1M)
docker pull ghcr.io/drowzeys/vllm-dspark-nvfp4-stage-c:gb10
docker tag ghcr.io/drowzeys/vllm-dspark-nvfp4-stage-c:gb10 \
vllm-dspark-runtime:dspark-nvfp4-stage-c
hf download drowzeys/DeepSeek-V4-Flash-DSpark-Abliterated-Uncensored-v1.1-alpha-Mida-Brikie-Mida-Brikie \
--local-dir ~/models/dsv4-flash-dspark-abliterated-mida
# rank1 then rank0 — see GitHub scripts/serve-v1.1-alpha.sh
MODELDIR=~/models/dsv4-flash-dspark-abliterated-mida \
CTX=1048576 SEQS=4 UTIL=0.85 bash scripts/serve-v1.1-alpha.sh 1
MODELDIR=~/models/dsv4-flash-dspark-abliterated-mida \
CTX=1048576 SEQS=4 UTIL=0.85 bash scripts/serve-v1.1-alpha.sh 0
Disclaimer
Research release. Outputs may include content stock models refuse. Do not deploy without your own safety layer. No liability for misuse.