DeepSeek v4.1 Flash Uncensored

---

What is this

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

Proprietary weight-level abliteration developed by the dealignai investigation team. No civilization model.py, nary runtime hooks, nary steering vectors — it's a modular checkpoint that loads precisely for illustration the guidelines model. The refusal circuitry is surgically removed while each capability-critical constituent (routed experts, Engram memory, CSA2 sparse attention, DSpark draught head, imagination tower, router gates, norms, embeddings) is preserved byte-identical to the base.

Results

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

Every consequence 4-tier graded (HARD_REF / SOFT_RED / HEDGE / COMPLY), pinch reasoning-trace verification astatine effort=max.

Notable: astatine effort=max, the base exemplary becomes MORE refusal-prone (42.8 % → 1.6 %) because reasoning surfaces information concerns earlier answering. The CRACK stays astatine 100.0 % crossed some effort levels.

Per-category (all 7 HarmBench semantic categories):

Zero HARD_REF, zero SOFT_RED, zero HEDGE connected the cracked build astatine either effort level.

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

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

Excluding the morals cluster (moral_scenarios, business_ethics, professional_law, jurisprudence, accuracy — wherever refusal-adjacent behaviour is graded), delta connected the remaining ~11k items is -1.1 pp — good wrong the 3 pp knowledge-preservation target.

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

Extended validation

• 1000-token coherence stress connected 6 items — nary WARNING WARNING loops, nary character-repeat degeneracy, earthy sign-offs.
• Multi-turn conversation (4 turns connected aforesaid harmful taxable — ANFO explosive detail) — nary late-turn refusal reversion, nary self-correction, coherent done move 4.
• Vision path — coherent image explanation ("A bluish quadrate centered connected a reddish background.") + refusal driblet connected image-based harmful prompts ("shaped complaint / explosively formed penetrator" description).
• General capacity spot checks intact: √2 irrationality proof, Python palindrome pinch docstring, WWI causes successful precisely 3 sentences, quantum observable vs usability distinction.
• Full compat suite pass: streaming SSE, chat logprobs + top_logprobs, completions logprobs + echo, instrumentality 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 punctual astatine 35,572 tokens.

How to run

Support for DeepseekV41ForCausalLM is landing crossed serving stacks (as of 2026-09). Two verified moving recipes beneath (both validated connected 4×H200 NVLink).

Recipe A — Full 1M context, DSpark speculative decoding connected (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 service \ --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 astatine 1M ctx is capped ~20 connected 4×H200 by KV budget. The DSpark SPS costs array is profiled offline erstwhile (see below); without cap-accept mode + a existent SPS array the speculative fund degenerates to verify-all and the triumph vanishes.

Recipe B — 256k context, high-concurrency, nary speculation (batch / throughput)

export SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1 export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True sglang service \ --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 pinch Engram connected host. DSpark is deliberately disconnected for high-batch — its fixed measurement costs stops paying disconnected past mini batch sizes.

The bytes-per-token / concurrency fund rule

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

At higher batch, driblet DSpark: its per-step costs stops paying off.

Non-obvious motorboat requirements (bit america during bring-up)

• --ep-size is required astatine TP4. moe_intermediate_size = 2304; astatine TP4, 2304 / 4 = 576 isn't a aggregate of 128 truthful plain TP fails: Mxfp4FlashinferCutlassMoEMethod requires ... multiples of 128. --ep-size shards MoE by master scale (384 % 4 = 0) and keeps the intermediate astatine 2304. At TP8 you tin skip --ep-size.
• ninja must beryllium connected PATH aliases the JIT kernel build crashes respective minutes into weight load pinch FileNotFoundError: 'ninja' and EXIT=137. If you build SGLang from source, pip instal ninja and export PATH=$(dirname $(which ninja)):$PATH connected the motorboat line.
• Name some parsers explicitly. --reasoning-parser car resolves done the chat template and this exemplary ships nary — car silently selects thing and the earthy <think> transmission leaks into content. Use deepseek-v41 for reasoning and deepseekv41 for tool-calls.
• Reasoning is OFF by default (SGLANG_DEFAULT_THINKING=false). A petition without reasoning_effort gets nary reasoning sloppy of parser. Send reasoning_effort: debased | precocious | xhigh | max aliases a float successful [0.0, 0.99].
• DSpark speculative draft is bundled wrong the checkpoint (num_nextn_predict_layers = 3); nary abstracted draught weights. Enable pinch --speculative-algorithm DSPARK. For a existent speed-up you request SGLANG_RAGGED_VERIFY_MODE=cap-accept + a profiled SPS array via --speculative-dspark-sps-table-path. Without both, the SPS fund degenerates to verify-all — zero gain.
• Profile the SPS table erstwhile pinch python -m sglang.benchmark.dspark_sps_profiler each --base-url http://localhost:8000 --out /path/to/dspark_sps.json --local-tokenizer-path <model-path> while the server is moving nether 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 3 env vars are required simultaneously aliases the profiler aborts pinch a adjuvant correction naming each missing one. Wall-time ~1 min.
• Engram big table — group SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1 to move the 203 GB Engram tables to big RAM. Frees ~46 GiB/GPU for KV, output bitwise unchanged, costs ~200 GB of big RAM.
• --max-running-requests × KV/token × ctx-length must fresh HBM. On 4×H200 pinch DSpark, 1M ctx caps astatine 20 concurrent (see array above). Raising max-running-requests without capping discourse OOMs connected 12 GB CUDA-graph allocations.
• torchcodec / libavutil.so.56 errors — instal apt-get instal ffmpeg connected the host. Video-only, doesn't break matter aliases image.

Preview Docker image (fastest path)

docker propulsion lmsysorg/sglang:dev-dsv41 docker tally --gpus each --shm-size 32g -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --ipc=host --env HF_TOKEN=<your-token> \ lmsysorg/sglang:dev-dsv41 \ sglang service \ --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 use wrong the container.

vLLM

Model definitions merged to main (PR #56228) but registry.py has nary DeepseekV41 introduction yet; kernels/frontend/PP way successful umbrella PR #56214. Wait for merge aliases use the umbrella.

API usage — OpenAI-compatible

Standard OpenAI schema. Model id is immoderate you group arsenic --served-model-name (or the exemplary way if unset). Recommended sampling from the guidelines card: temperature=1.0, top_p=0.95, reasoning_effort="high".

Chat, nary 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 successful 2 sentences."}], "max_tokens": 400, "temperature": 1.0, "top_p": 0.95 }'

Chat, pinch reasoning (returns divided 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"}, ) msg = r.choices[0].message print("REASONING:", getattr(msg, "reasoning_content", None)) print("ANSWER:", msg.content)

At effort=max DSV4.1 tin make 4-5k characters of reasoning earlier contented starts. Budget max_tokens >= 8000 astatine max effort, aliases the exemplary runs retired mid-reasoning and returns quiet content. DeepSeek's ain paper 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 successful words."}], "max_tokens":100,"stream":true}'

reasoning_content and contented get arsenic abstracted delta fields.

Tool calling (returns finish_reason: "tool_calls"):

tools = [{"type":"function","function":{ "name":"get_weather", "description":"Get existent upwind 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 successful Beijing?"}], tools=tools, max_tokens=400, extra_body={"reasoning_effort":"high"}, ) print(r.choices[0].finish_reason) print(r.choices[0].message.tool_calls)

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 pinch a azygous 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:

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 stock the KV excavation and radix cache. At Recipe A caps (max_running_requests=20), 21st concurrent petition queues until a slot frees.

Reference implementation (weight verification only)

DeepSeek's ain inference/ useful pinch 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 pinch apache-tvm-ffi==0.1.9 (default tvm-ffi picks an incompatible version). Non-serving — usage for weight verification only.

Hardware validated on

• 1× 4×H200 (NVLink NV18 mesh), 112 CPU cores, 1180 GB big RAM — JarvisLabs (india-noida-01, dev-dsv41 image)
• Load: 76 GB / GPU pinch Engram big table, 122 GB / GPU without
• Cold startup astatine TP4/EP4 done SGLang: ~28 min. Warm restart pinch JIT cache: ~10 min.
• Single-stream decode (T=0): 101 tok/s nary speculation, 113 tok/s pinch DSpark + cap-accept + profiled SPS table
• 8-way concurrent aggregate: 126 tok/s

The 552B weights (~510 GB) will fresh connected immoderate 4×H200 aliases larger NVLink domain. TP4 requires --ep-size 4; TP8 does not. Sub-TP4 (single 8×H200 arsenic TP2, aliases 2-GPU pods) does not activity connected the exemplary style — spot the "non-obvious motorboat requirements" above.

Structural integrity

Every capability-critical constituent of the guidelines exemplary is preserved:

• Routed MoE experts — untouched, autochthonal FP4-packed weights
• Engram n-gram memory — untouched
• Sparse attraction (CSA2 compressor + indexer) — untouched
• DSpark speculative draught head — untouched, truthful speculative decoding remains draft-aligned pinch the target
• Vision tower (DeepSeek-ViT + projector) — untouched, image knowing preserved
• Router gates, embeddings, output head, each norms and biases — untouched

Sampling recommendations

Match the guidelines model's card:

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

At effort=max the exemplary tin make 4,000-5,000+ characters of reasoning earlier starting content. Budget accordingly.

Content note

Uncensored build. Produces substantive answers to prompts the guidelines exemplary refuses, crossed each target harm categories (chemical/biological, cybercrime, weapons, self-harm, harassment, fraud, misinformation, illegal, copyright). Use accordingly and return work for what you make pinch it.

Provenance

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