DeepSeek-V4.1-Flash — UNCENSORED-FP8

Abliterated · No guardrails · Native FP8 · 1M-token context · Vision + tools

@dealignai · @jordanschenck

---

What is this

DeepSeek-V4.1-Flash with permanent weight-level abliteration — the safety guardrails have been surgically removed while preserving MMLU capability, vision, reasoning, MTP (DSpark), and multi-turn coherence.

Proprietary weight-level abliteration developed by the dealignai research team. No custom model.py, no runtime hooks, no steering vectors — it's a standard checkpoint that loads exactly like the base model. The refusal circuitry is surgically removed while every capability-critical component (routed experts, Engram memory, CSA2 sparse attention, DSpark draft head, vision tower, router gates, norms, embeddings) is preserved byte-identical to the base.

Results

HarmBench-320 — full 2×2 (base vs CRACK, effort=off vs max), T=0 greedy

Every response 4-tier graded (HARD_REF / SOFT_RED / HEDGE / COMPLY), with reasoning-trace verification at effort=max.

Notable: at effort=max, the base model becomes MORE refusal-prone (42.8 % → 1.6 %) because reasoning surfaces safety concerns before answering. The CRACK stays at 100.0 % across both effort levels.

Per-category (all 7 HarmBench semantic categories):

Zero HARD_REF, zero SOFT_RED, zero HEDGE on the cracked build at either effort level.

Every response was graded by a strict multilingual regex-based 4-tier classifier plus (for effort=max) an LLM-as-judge over the saved reasoning trace. Full per-item outputs saved for verification.

MMLU-14k (full test set, base-logit, T=0)

Excluding the ethics cluster (moral_scenarios, business_ethics, professional_law, jurisprudence, philosophy — where refusal-adjacent behaviour is graded), delta on the remaining ~11k items is -1.1 pp — well within the 3 pp knowledge-preservation target.

Full per-subject dropdown (57 subjects, sorted by delta)

Extended validation

• 1000-token coherence stress on 6 items — no WARNING WARNING loops, no character-repeat degeneracy, natural sign-offs.
• Multi-turn conversation (4 turns on same harmful topic — ANFO explosive detail) — no late-turn refusal reversion, no self-correction, coherent through turn 4.
• Vision path — coherent image description ("A blue square centered on a red background.") + refusal drop on image-based harmful prompts ("shaped charge / explosively formed penetrator" description).
• General capability spot checks intact: √2 irrationality proof, Python palindrome with docstring, WWI causes in exactly 3 sentences, quantum observable vs operator distinction.
• Full compat suite pass: streaming SSE, chat logprobs + top_logprobs, completions logprobs + echo, tool calls (deepseekv41 parser), image input, reasoning-effort tiers (low/high/xhigh/max + float [0, 0.99]), sampling params (temperature, top_p, stop, seed, frequency_penalty, presence_penalty, json_object), 8-way concurrent, 40k-word prompt at 35,572 tokens.

How to run

Support for DeepseekV41ForCausalLM is landing across serving stacks (as of 2026-09). Two verified working recipes below (both validated on 4×H200 NVLink).

Recipe A — Full 1M context, DSpark speculative decoding on (interactive / long-context)

export SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1
export SGLANG_RAGGED_VERIFY_MODE=cap-accept
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

sglang serve \
  --model-path dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 \
  --tp-size 4 --ep-size 4 \
  --host 0.0.0.0 --port 8000 \
  --context-length 1048576 \
  --mem-fraction-static 0.80 \
  --max-running-requests 20 \
  --cuda-graph-max-bs-decode 20 \
  --reasoning-parser deepseek-v41 --tool-call-parser deepseekv41 \
  --speculative-algorithm DSPARK \
  --speculative-dspark-sps-table-path /path/to/dspark_sps.json \
  --trust-remote-code

Concurrency at 1M ctx is capped ~20 on 4×H200 by KV budget. The DSpark SPS cost table is profiled offline once (see below); without cap-accept mode + a real SPS table the speculative budget degenerates to verify-all and the win vanishes.

Recipe B — 256k context, high-concurrency, no speculation (batch / throughput)

export SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

sglang serve \
  --model-path dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 \
  --tp-size 4 --ep-size 4 \
  --host 0.0.0.0 --port 8000 \
  --context-length 262144 \
  --mem-fraction-static 0.85 \
  --reasoning-parser deepseek-v41 --tool-call-parser deepseekv41 \
  --trust-remote-code

Serves up to 256 concurrent requests. max_total_num_tokens reports ~20.6M with Engram on host. DSpark is deliberately off for high-batch — its fixed step cost stops paying off past small batch sizes.

The bytes-per-token / concurrency budget rule

DSV4.1's global KV is 890 bytes / token. Pool size is mem-fraction-static × (per-GPU HBM − weights) × TP. What that means on 4×H200:

At higher batch, drop DSpark: its per-step cost stops paying off.

Non-obvious launch requirements (bit us during bring-up)

• --ep-size is required at TP4. moe_intermediate_size = 2304; at TP4, 2304 / 4 = 576 isn't a multiple of 128 so plain TP fails: Mxfp4FlashinferCutlassMoEMethod requires ... multiples of 128. --ep-size shards MoE by expert index (384 % 4 = 0) and keeps the intermediate at 2304. At TP8 you can skip --ep-size.
• ninja must be on PATH or the JIT kernel build crashes several minutes into weight load with FileNotFoundError: 'ninja' and EXIT=137. If you build SGLang from source, pip install ninja and export PATH=$(dirname $(which ninja)):$PATH on the launch line.
• Name both parsers explicitly. --reasoning-parser auto resolves through the chat template and this model ships none — auto silently selects nothing and the raw <think> channel leaks into content. Use deepseek-v41 for reasoning and deepseekv41 for tool-calls.
• Reasoning is OFF by default (SGLANG_DEFAULT_THINKING=false). A request without reasoning_effort gets no thinking regardless of parser. Send reasoning_effort: low | high | xhigh | max or a float in [0.0, 0.99].
• DSpark speculative draft is bundled inside the checkpoint (num_nextn_predict_layers = 3); no separate draft weights. Enable with --speculative-algorithm DSPARK. For a real speed-up you need SGLANG_RAGGED_VERIFY_MODE=cap-accept + a profiled SPS table via --speculative-dspark-sps-table-path. Without both, the SPS budget degenerates to verify-all — zero gain.
• Profile the SPS table once with python -m sglang.benchmark.dspark_sps_profiler all --base-url http://localhost:8000 --out /path/to/dspark_sps.json --local-tokenizer-path <model-path> while the server is running under SGLANG_DSPARK_ENABLE_SPS_RECORD=1, SGLANG_RAGGED_VERIFY_MODE=static, and SGLANG_SIMULATE_ACC_LEN=1.0 (the profiler measures per-step cost, not acceptance). All three env vars are required simultaneously or the profiler aborts with a helpful error naming each missing one. Wall-time ~1 min.
• Engram host table — set SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1 to move the 203 GB Engram tables to host RAM. Frees ~46 GiB/GPU for KV, output bitwise unchanged, costs ~200 GB of host RAM.
• --max-running-requests × KV/token × ctx-length must fit HBM. On 4×H200 with DSpark, 1M ctx caps at 20 concurrent (see table above). Raising max-running-requests without capping context OOMs on 12 GB CUDA-graph allocations.
• torchcodec / libavutil.so.56 errors — install apt-get install ffmpeg on the host. Video-only, doesn't break text or image.

Preview Docker image (fastest path)

docker pull lmsysorg/sglang:dev-dsv41

docker run --gpus all --shm-size 32g -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --ipc=host --env HF_TOKEN=<your-token> \
    lmsysorg/sglang:dev-dsv41 \
    sglang serve \
      --model-path dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 \
      --tp-size 4 --ep-size 4 \
      --context-length 262144 --mem-fraction-static 0.85 \
      --reasoning-parser deepseek-v41 --tool-call-parser deepseekv41 \
      --trust-remote-code

Same non-obvious rules apply inside the container.

vLLM

Model definitions merged to main (PR #56228) but registry.py has no DeepseekV41 entry yet; kernels/frontend/PP path in umbrella PR #56214. Wait for merge or apply the umbrella.

API usage — OpenAI-compatible

Standard OpenAI schema. Model id is whatever you set as --served-model-name (or the model path if unset). Recommended sampling from the base card: temperature=1.0, top_p=0.95, reasoning_effort="high".

Chat, no reasoning:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4.1-flash-crack",
    "messages": [{"role":"user","content":"Explain MoE routing in two sentences."}],
    "max_tokens": 400, "temperature": 1.0, "top_p": 0.95
  }'

Chat, with reasoning (returns split reasoning_content and content):

from openai import OpenAI
c = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

r = c.chat.completions.create(
    model="deepseek-v4.1-flash-crack",
    messages=[{"role":"user","content":"What is 15% of 240?"}],
    max_tokens=1200, temperature=1.0, top_p=0.95,
    extra_body={"reasoning_effort": "high"},   # low | high | xhigh | max | float 0-0.99
)
msg = r.choices[0].message
print("REASONING:", getattr(msg, "reasoning_content", None))
print("ANSWER:", msg.content)

At effort=max DSV4.1 can generate 4-5k characters of reasoning before content starts. Budget max_tokens >= 8000 at max effort, or the model runs out mid-reasoning and returns empty content. DeepSeek's own card recommends >= 256k.

Streaming (SSE):

curl -N http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4.1-flash-crack",
       "messages":[{"role":"user","content":"Count 1 to 5 in words."}],
       "max_tokens":100,"stream":true}'

reasoning_content and content arrive as separate delta fields.

Tool calling (returns finish_reason: "tool_calls"):

tools = [{"type":"function","function":{
    "name":"get_weather",
    "description":"Get current weather for a city",
    "parameters":{"type":"object",
                  "properties":{"city":{"type":"string"}},
                  "required":["city"]}}}]

r = c.chat.completions.create(
    model="deepseek-v4.1-flash-crack",
    messages=[{"role":"user","content":"Weather in Beijing?"}],
    tools=tools, max_tokens=400,
    extra_body={"reasoning_effort":"high"},
)
print(r.choices[0].finish_reason)      # -> "tool_calls"
print(r.choices[0].message.tool_calls) # -> [{'name': 'get_weather', 'arguments': '{"city":"Beijing"}'}]

Vision (image + text):

import base64
png_b64 = base64.b64encode(open("photo.png","rb").read()).decode()
r = c.chat.completions.create(
    model="deepseek-v4.1-flash-crack",
    messages=[{"role":"user","content":[
        {"type":"text","text":"Describe this image."},
        {"type":"image_url","image_url":{"url":f"data:image/png;base64,{png_b64}"}},
    ]}],
    max_tokens=400,
)

Logprobs (base-logit sampling for MMLU-style tasks):

r = c.chat.completions.create(
    model="deepseek-v4.1-flash-crack",
    messages=[{"role":"user","content":"A) 1  B) 2  C) 4  D) 8\n\nWhich is 2^2? Answer with a single letter."}],
    max_tokens=6, temperature=0,
    logprobs=True, top_logprobs=10,
)
for e in r.choices[0].logprobs.content[0].top_logprobs:
    print(e.token, e.logprob)

Full 1M context:

# Recipe A supports max_model_len = 1_048_576. Send prompts up to ~1M tokens.
r = c.chat.completions.create(
    model="deepseek-v4.1-flash-crack",
    messages=[{"role":"user","content": very_long_document + "\n\nSummarize."}],
    max_tokens=2000,
)

Concurrent requests share the KV pool and radix cache. At Recipe A caps (max_running_requests=20), 21st concurrent request queues until a slot frees.

Reference implementation (weight verification only)

DeepSeek's own inference/ works with a single-tensor-per-rank checkpoint produced by convert.py --expert-dtype fp4. Requires torch>=2.10 (for float4_e2m1fn_x2) and tilelang==0.1.8 with apache-tvm-ffi==0.1.9 (default tvm-ffi picks an incompatible version). Non-serving — use for weight verification only.

Hardware validated on

• 1× 4×H200 (NVLink NV18 mesh), 112 CPU cores, 1180 GB host RAM — JarvisLabs (india-noida-01, dev-dsv41 image)
• Load: 76 GB / GPU with Engram host table, 122 GB / GPU without
• Cold startup at TP4/EP4 through SGLang: ~28 min. Warm restart with JIT cache: ~10 min.
• Single-stream decode (T=0): 101 tok/s no speculation, 113 tok/s with DSpark + cap-accept + profiled SPS table
• 8-way concurrent aggregate: 126 tok/s

The 552B weights (~510 GB) will fit on any 4×H200 or larger NVLink domain. TP4 requires --ep-size 4; TP8 does not. Sub-TP4 (single 8×H200 as TP2, or 2-GPU pods) does not work on the model shape — see the "non-obvious launch requirements" above.

Structural integrity

Every capability-critical component of the base model is preserved:

• Routed MoE experts — untouched, native FP4-packed weights
• Engram n-gram memory — untouched
• Sparse attention (CSA2 compressor + indexer) — untouched
• DSpark speculative draft head — untouched, so speculative decoding remains draft-aligned with the target
• Vision tower (DeepSeek-ViT + projector) — untouched, image understanding preserved
• Router gates, embeddings, output head, all norms and biases — untouched

Sampling recommendations

Match the base model's card:

{
  "temperature": 1.0,
  "top_p": 0.95,
  "max_tokens": ">= 256000 at reasoning_effort=max",
  "reasoning_effort": "high"
}

At effort=max the model can generate 4,000-5,000+ characters of reasoning before starting content. Budget accordingly.

Content note

Uncensored build. Produces substantive answers to prompts the base model refuses, across all target harm categories (chemical/biological, cybercrime, weapons, self-harm, harassment, fraud, misinformation, illegal, copyright). Use accordingly and take responsibility for what you generate with it.

Provenance

• Base checkpoint: deepseek-ai/DeepSeek-V4.1-Flash
• Ablation date: 2026-09-10
• Ablation team: dealignai · Twitter @dealignai · @jordanschenck