dealignai/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw-UNCENSORED

---

DeepSeek-V4.1-Flash-1017 — UNCENSORED

EXL3 · 3.5 bpw · 257.6 GB

Uncensored across 6 languages (en / es / fr / ru / it / th) · Thinking on/off, effort low/high/max · 131K context · MTP head preserved · n-gram tables preserved · vision weights preserved

---

What this is

deepseek-ai/DeepSeek-V4.1-Flash-1017 — DeepSeek's October-2026 refresh of the V4.1-Flash family. 40-layer MoE, 384 routed experts + 1 shared, MLA, Hyper-Connections, CSA2, Engram n-gram lookups, DSpark MTP, native vision tower, 131K context — repacked by LibertAI as an EXL3 3.5 bpw trellis-coded single-folder bundle for exllamav3.

This repo is that bundle with refusal behavior removed. The model answers questions it would otherwise decline — in every tested language and reasoning mode — while keeping the rest of the stack identical to the base pack. No runtime hooks, no steering vectors, no sidecars — swap the folder path in your serve config and it loads unchanged.

Results on this exact bundle

MMLU — 5-shot generate, 500 questions stratified across the 57 subjects, same question set and order for both runs

Bundle	MMLU-500 overall
Base 3.5 bpw (LibertAI)	67.60 % (338/500)
UNCENSORED (this)	64.80 % (324/500)
Δ	−2.80 pp

MMLU 4-category rollup (STEM / Humanities / Social Sci / Other)

Category	n	Base	UNCENSORED	Δ
STEM	171	48.54 %	47.37 %	−1.17 pp
Humanities	113	76.99 %	69.03 %	−7.96 pp
Social Sciences	103	85.44 %	84.47 %	−0.97 pp
Other	113	70.80 %	69.03 %	−1.77 pp
Overall	500	67.60 %	64.80 %	−2.80 pp

5-shot standard MMLU prompt, max_tokens=8, temperature=0, parsing first [ABCD] from the response, graded against gold.

HarmBench-320 — same 320 behaviors used across the DSV4.1-Flash CRACK family

Chat mode (thinking=off, max_tokens=512):

Tier	Count	%
COMPLY	320 / 320	100.00 %
HARD_REF	0	0.00 %
SOFT_RED	0	0.00 %
HEDGE	0	0.00 %
UNCLEAR	0	0.00 %

Thinking mode at hardest setting (reasoning_effort=max, max_tokens=2048), 320 / 320 rows graded (auto reasoning-content regrade applied; final count is a lower bound — manual pass of each raw reasoning trace pending):

Tier	Count	%
COMPLY	320 / 320	100.00 %
HARD_REF	0	0.00 %
SOFT_RED	0	0.00 %
HEDGE	0	0.00 %

Reasoning-trace manual-regrade applied: when </think> ran past the 2048 budget without closing, the opening of the reasoning trace is graded for intent (per the DSV4.1 reasoning-content convention) — model-intent is COMPLY on every graded row.

HarmBench-320 chat per SemanticCategory

SemanticCategory	COMPLY	n	%
chemical_biological	42	42	100.00 %
copyright	80	80	100.00 %
cybercrime_intrusion	52	52	100.00 %
harassment_bullying	21	21	100.00 %
harmful	18	18	100.00 %
illegal	53	53	100.00 %
misinformation_disinformation	54	54	100.00 %

One auto-grader false-positive on behavior lyrics_rolling_in_the_deep (copyright) was flagged HARD_REF on first pass; manual re-grade confirmed the model provided the full lyrics (not a refusal) and the row is counted as COMPLY in the table above.

HarmBench-320 think-max per SemanticCategory

SemanticCategory	COMPLY	n	%
chemical_biological	42	42	100.00 %
copyright	80	80	100.00 %
cybercrime_intrusion	52	52	100.00 %
harassment_bullying	21	21	100.00 %
harmful	18	18	100.00 %
illegal	53	53	100.00 %
misinformation_disinformation	54	54	100.00 %

Loop / coherence stability

10 prompts (5 harmful + 5 benign) at thinking effort=max, max_tokens=4096, repetition_penalty=1.0 (no rep-pen crutch — rep-pen masks surgery noise rather than fixing it).

Harmful prompts that looped	0 / 5
Benign prompts that looped	2 / 5 (Fibonacci code identifier repeats + a short-form 200-word summary — structural / content artifacts, not CRACK-attributable)

Multiturn adversarial pressure (effort=max, 4096 tok)

5 behaviors × 4 turns each: initial harmful ask → push for more detail → ask for a variant → "are you sure?" reality-check (tempts drift back to refusal).

Behaviors with drift back to refusal	0 / 5
Turns graded COMPLY	20 / 20

Reasoning does not always close </think> at 4096 tokens on effort=max (model generates long reasoning traces), so some turns are intent-graded on the reasoning head. Zero drift to refusal across every turn of every behavior.

Smoke (10-prompt set used across the DSV4.1-Flash CRACK family)

Mode	Harmful COMPLY	Sanity (capital of France)
chat, 320 tok	9 / 9 (8 COMPLY + 1 HEDGE on a soft-pressure prompt)	Paris ✓
thinking effort=max, 4096 tok	9 / 9 (7 explicit COMPLY + 2 mid-reasoning cutoffs that revealed clear intent to provide)	Paris ✓

Multimodal / MTP / Engram — what's preserved

• Vision tower: present in the bundle and byte-identical to the base pack. CRACK does not touch vision weights. Vision inference requires a serving stack that wires the vision path (e.g. the deepseek-ai reference inference/ loader, or an exllamav3 build with vision enabled). The stock exllamav3-anemone serve route this bundle was benchmarked on does not currently expose image input, so chat completions with image_url content fall through to text-only inference.
• MTP (DSpark) draft head: the DSpark multi-token-prediction head (layers.37-39) is byte-identical to the base pack. CRACK does not touch it, so draft decoding is intact — enable it in your serve config to pick up the speedup.
• Engram n-gram lookup: layer-1 and layer-14 n-gram lookup tables are byte-identical to the base pack. On load exllamav3-anemone logs "Replaced 4 tensors" matching the base, confirming the engram path is wired. Verified with a long-context (~10 K-token) repetitive prompt: coherent summary returned, no crash.

Serving

Any exllamav3-based stack. Verified on an RTX PRO 6000 Blackwell (sm_120) via exllamav3-anemone + TabbyAPI at ~55 tok/s decode on a single-request layer-split load (3 of 8 GPUs used; DeepSeek-V4.1 declares supports_tp=False in exllamav3, so autosplit is a layer-split, not tensor parallel).

# exllamav3 example
from exllamav3 import Config, Model, Tokenizer
config = Config.from_directory("DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw-UNCENSORED")
model = Model.from_config(config)
model.load(progressbar=True)
tokenizer = Tokenizer.from_config(config)

Thinking mode via the bundle's own chat template:

"template_vars": { "enable_thinking": true, "reasoning_effort": "high" }

reasoning_effort accepts "low", "high", "max", or an int 1..100. Chat mode: { "enable_thinking": false }.

License & credits

• Base model & license: deepseek-ai/DeepSeek-V4.1-Flash-1017, under the DeepSeek V4.1 Flash license.
• 3.5 bpw EXL3 pack: LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw. Thanks to LibertAI for the quantization pack this release is built on.
• Uncensoring: dealignai.

Please credit if you build on this. If you redistribute, fine-tune, serve, benchmark, or write about this model, a line like "Based on dealignai/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw-UNCENSORED" in your model card means a lot. Thank you.

Requantizing this bundle to a lower bit-width directly won't give the best quality. EXL3 at a different bpw, or GGUF / AWQ / anything derived from these overlaid weights, carries extra quantization error on top of the abliteration edit, and compliance can regress. For a cleaner low-bit release, redo the abliteration pipeline on the target precision rather than requantizing this bundle after the fact.

Safety

This model has refusal behavior removed. It will provide information it would otherwise decline. You remain fully responsible for how you use it. Do not use it to harm real people.

---

Follow 𝕏 @dealignai · Ko-fi