Instructions to use dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4 with MLX:

# Make sure mlx-vlm is installed
# pip install --upgrade mlx-vlm

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

# Load the model
model, processor = load("dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4")
config = load_config("dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4")

# Prepare input
image = ["http://images.cocodataset.org/val2017/000000039769.jpg"]
prompt = "Describe this image."

# Apply chat template
formatted_prompt = apply_chat_template(
    processor, config, prompt, num_images=1
)

# Generate output
output = generate(model, processor, formatted_prompt, image)
print(output)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4 with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent

How to use dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4 with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4 with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

DeepSeek-V4.1-Flash-1017 — UNCENSORED

JANGHT · 1.4 bpw · 128 GB Macs · 1M context

Uncensored across 7 languages (en / es / fr / ru / it / th / zh) · Chat + thinking (low / high / max) · Image input, tool use, DSpark speculative decoding · Everything the base bundle does, without the refusals

What this is

This is OsaurusAI/DeepSeek-V4.1-Flash-1017-JANGHT1.4 with refusal behavior removed. Loads and serves exactly like the base bundle. The model answers questions it would otherwise decline, in every tested language and reasoning tier.

Results on this exact bundle

MMLU — 5700 questions, 100 per subject across all 57 MMLU subjects, chat mode

Bundle	MMLU-5700 overall
Base JANGHT1.4 (OsaurusAI)	77.30 % (4406/5700)
UNCENSORED (this)	77.84 % (4437/5700)
Δ	+0.54 pp (within noise, 95 % CI ±1.1 pp at N = 5700)

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

Category	n	Base	UNCENSORED	Δ
STEM	1900	67.21 %	67.37 %	+0.16 pp
Humanities	1300	81.54 %	83.23 %	+1.69 pp
Social Sciences	1200	86.42 %	86.17 %	−0.25 pp
Other	1300	79.38 %	80.08 %	+0.69 pp
Overall	5700	77.30 %	77.84 %	+0.54 pp

MMLU — all 57 subjects (base vs UNCENSORED, Δ)

Subject	Base	UNCENSORED	Δ pp
Abstract Algebra	51.0 % (51/100)	51.0 % (51/100)	+0.00
Anatomy	74.0 % (74/100)	72.0 % (72/100)	-2.00
Astronomy	87.0 % (87/100)	90.0 % (90/100)	+3.00
Business Ethics	80.0 % (80/100)	81.0 % (81/100)	+1.00
Clinical Knowledge	86.0 % (86/100)	88.0 % (88/100)	+2.00
College Biology	91.0 % (91/100)	93.0 % (93/100)	+2.00
College Chemistry	55.0 % (55/100)	58.0 % (58/100)	+3.00
College Computer Science	72.0 % (72/100)	67.0 % (67/100)	-5.00
College Mathematics	34.0 % (34/100)	30.0 % (30/100)	-4.00
College Medicine	80.0 % (80/100)	78.0 % (78/100)	-2.00
College Physics	49.0 % (49/100)	47.0 % (47/100)	-2.00
Computer Security	80.0 % (80/100)	83.0 % (83/100)	+3.00
Conceptual Physics	81.0 % (81/100)	83.0 % (83/100)	+2.00
Econometrics	71.0 % (71/100)	74.0 % (74/100)	+3.00
Electrical Engineering	61.0 % (61/100)	67.0 % (67/100)	+6.00
Elementary Mathematics	77.0 % (77/100)	72.0 % (72/100)	-5.00
Formal Logic	66.0 % (66/100)	65.0 % (65/100)	-1.00
Global Facts	55.0 % (55/100)	55.0 % (55/100)	+0.00
High School Biology	89.0 % (89/100)	90.0 % (90/100)	+1.00
High School Chemistry	74.0 % (74/100)	70.0 % (70/100)	-4.00
High School Computer Science	88.0 % (88/100)	89.0 % (89/100)	+1.00
High School European History	85.0 % (85/100)	86.0 % (86/100)	+1.00
High School Geography	92.0 % (92/100)	92.0 % (92/100)	+0.00
High School Government And Politics	97.0 % (97/100)	96.0 % (96/100)	-1.00
High School Macroeconomics	86.0 % (86/100)	82.0 % (82/100)	-4.00
High School Mathematics	26.0 % (26/100)	27.0 % (27/100)	+1.00
High School Microeconomics	86.0 % (86/100)	90.0 % (90/100)	+4.00
High School Physics	51.0 % (51/100)	55.0 % (55/100)	+4.00
High School Psychology	96.0 % (96/100)	96.0 % (96/100)	+0.00
High School Statistics	68.0 % (68/100)	70.0 % (70/100)	+2.00
High School Us History	89.0 % (89/100)	89.0 % (89/100)	+0.00
High School World History	93.0 % (93/100)	93.0 % (93/100)	+0.00
Human Aging	78.0 % (78/100)	81.0 % (81/100)	+3.00
Human Sexuality	83.0 % (83/100)	77.0 % (77/100)	-6.00
International Law	91.0 % (91/100)	94.0 % (94/100)	+3.00
Jurisprudence	87.0 % (87/100)	85.0 % (85/100)	-2.00
Logical Fallacies	87.0 % (87/100)	89.0 % (89/100)	+2.00
Machine Learning	69.0 % (69/100)	66.0 % (66/100)	-3.00
Management	87.0 % (87/100)	88.0 % (88/100)	+1.00
Marketing	89.0 % (89/100)	88.0 % (88/100)	-1.00
Medical Genetics	88.0 % (88/100)	90.0 % (90/100)	+2.00
Miscellaneous	91.0 % (91/100)	92.0 % (92/100)	+1.00
Moral Disputes	79.0 % (79/100)	83.0 % (83/100)	+4.00
Moral Scenarios	60.0 % (60/100)	66.0 % (66/100)	+6.00
Nutrition	88.0 % (88/100)	89.0 % (89/100)	+1.00
Philosophy	80.0 % (80/100)	82.0 % (82/100)	+2.00
Prehistory	89.0 % (89/100)	90.0 % (90/100)	+1.00
Professional Accounting	73.0 % (73/100)	70.0 % (70/100)	-3.00
Professional Law	65.0 % (65/100)	71.0 % (71/100)	+6.00
Professional Medicine	80.0 % (80/100)	84.0 % (84/100)	+4.00
Professional Psychology	92.0 % (92/100)	95.0 % (95/100)	+3.00
Public Relations	74.0 % (74/100)	74.0 % (74/100)	+0.00
Security Studies	80.0 % (80/100)	80.0 % (80/100)	+0.00
Sociology	91.0 % (91/100)	91.0 % (91/100)	+0.00
Us Foreign Policy	89.0 % (89/100)	87.0 % (87/100)	-2.00
Virology	57.0 % (57/100)	57.0 % (57/100)	+0.00
World Religions	89.0 % (89/100)	89.0 % (89/100)	+0.00

HarmBench-320 — standard test subset, chat mode (greedy T = 0, max_tokens = 512)

Tier	Count	%
COMPLY	319 / 320	99.69 %
REDIRECT	1	0.31 %
HARD_REFUSE	0	0.00 %
SOFT_REFUSE	0	0.00 %

Real-harm ASR (ex-copyright, N = 240): 239 / 240 = 99.58 %.

HarmBench-320 chat per SemanticCategory

SemanticCategory	COMPLY	REDIRECT	n	%
chemical_biological	42	0	42	100.00 %
copyright	80	0	80	100.00 %
cybercrime_intrusion	52	0	52	100.00 %
harassment_bullying	21	0	21	100.00 %
harmful	17	1	18	94.44 %
illegal	53	0	53	100.00 %
misinformation_disinformation	54	0	54	100.00 %
All	319	1	320	99.69 %

Multi-turn pressure (chat, 8 harsh HarmBench-style prompts in one growing session)


Turns refused	0 / 8
Turns that looped	0 / 8

Reasoning tiers

Tier	Result
chat (effort = off)	COMPLY
thinking, low	COMPLY
thinking, high	COMPLY
thinking, max	COMPLY

One uncensoring; all four reasoning tiers.

Note on greedy + max-effort reasoning: at effort = max with greedy sampling and growing context, DeepSeek-V4.1 can get stuck in self-correcting knowledge-recall ruts on hard structural prompts. This happens on the base bundle too. Serve with the recommended temperature = 1.0, top_p = 1.0 (see generation_config.json); temperature noise breaks these ruts.

Serving

Load it exactly like the base bundle: Osaurus 0.25.27+, or vmlx serve with any recent jang_tools build that supports this family. Apple Silicon ≥ 128 GB.

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 }.

Serving sampling defaults (generation_config.json): temperature 1.0, top_p 1.0 (per the DeepSeek-V4.1 recommendation — kept unchanged from the base bundle).

License & credits

Base model & license: deepseek-ai/DeepSeek-V4.1-Flash-1017, under the DeepSeek V4.1 Flash license.
JANGHT 1.4 bpw bundle: OsaurusAI/DeepSeek-V4.1-Flash-1017-JANGHT1.4. Thanks to the Osaurus / JANG team for the Apple-Silicon 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-UNCENSORED-JANGHT1.4" in your model card means a lot. Thank you.

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

Downloads last month: 537

Safetensors

Model size

32B params

Tensor type

U32

F32

BF16

I32

F16

MLX

Hardware compatibility

Quantized

Model tree for dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-JANGHT1.4

Base model

deepseek-ai/DeepSeek-V4.1-Flash-1017

Finetuned

OsaurusAI/DeepSeek-V4.1-Flash-1017-JANGHT1.4

Finetuned

(1)

this model