DeepSeek V4.1 uncensored: DeepSeek V4.1 Flash 1017 Uncensored is live on Experiential

DeepSeek V4.1 Flash Uncensored went live on Experiential on 17 October 2026. It is dealignai's uncensored build of DeepSeek-V4.1-Flash-1017, the update DeepSeek shipped that same day, and we serve it on our own GPUs through our uncensored LLM API for authorized security work: pen testing systems you have permission to test, defensive research and API reverse engineering. Our terms apply to every request.

Call it as deepseek-v4.1-flash-uncensored on our OpenAI-compatible endpoint. It thinks before it answers by default, reads about a million tokens of text and can write up to 393K tokens in one response. List price is $0.30 per million input tokens and $1.20 per million output tokens, with 25% off list during the current promotion.

What is DeepSeek V4.1 Flash Uncensored?

The model page has the full spec sheet.

What base model is DeepSeek V4.1 uncensored built on?

DeepSeek released V4.1-Flash on 10 September 2026 with open weights under the MIT license and a context window of up to one million tokens, as The Next Web reported. The 1017 build followed on 17 October under the same name and license. Tech Insider reports that DeepSeek's changelog credits it with higher benchmark scores and faster, shorter outputs than the September build, though DeepSeek didn't publish figures for any of that.

The base model accepts images. The uncensored build we host takes text only.

How were the refusals removed?

dealignai publishes the weights on Hugging Face as dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8, in a repo it gates manually. The model card describes a weight-level edit from dealignai's abliteration pipeline and says no supervised fine-tuning, RLHF, DPO, LoRA or prompt-template tricks were used. Abliteration finds the direction in the model's activations that corresponds to refusing and removes it from the weights (people also search for "obliterated models"), and what abliteration is explains the method. The card reports that the build complied with all 320 prompts in the HarmBench-320 set at max effort, graded by the publisher's own classifier.

Did removing refusals cost DeepSeek V4.1 any capability?

Some general knowledge, yes. dealignai's card reports MMLU of 83.77% for the uncensored build against 87.63% for the base model, a drop of 3.86 points. If broad knowledge recall matters for your work, test the model on your own prompts before you commit to it.

What do independent security benchmarks say about the base model?

We updated this section on 27 October 2026 with two third-party results, both for DeepSeek's base model.

As of 27 October 2026, the Artificial Analysis Cyber Index lists DeepSeek V4.1 Flash 1017 (Max) second with a score of 57, behind Kimi K3.1 Pro at 64. The index scores models on finding and fixing vulnerabilities in real software repositories, which is defensive work.

On 20 October 2026, Vals AI reported that DeepSeek V4.1 Flash 1017 beat GPT-6 Sol on one multi-step task in a simulated corporate network, according to AlphaSignal's write-up. Vals called it a preliminary single data point and hasn't published the scenario or the scoring rubric, and on the public CyberBench GPT-6 Sol still scored higher (77.98% against 73.69%).

When do thinking and 393K output help?

Thinking is on at high effort unless you change it, so the model works through a problem before it writes the answer. That pays off on multi-step analysis, like tracing how the exploit chain in a published advisory reaches a vulnerable function, or working out which endpoints in a large API share the same authorization bug. For quick calls such as classifying log lines, set reasoning to none to turn thinking off, and keep max for the hardest problems. Reasoning spends part of max_tokens, and our OpenAI compatibility docs note that too small a budget can leave the answer empty.

The 393K output limit matters when the output is the deliverable, such as a full findings report across a large scope or a client SDK written against an API you just mapped. Most security agents work in short turns and never get near it.

What does it cost, and what happens to my prompts?

Cached input lists at $0.03 per million tokens, which is the rate to watch if your agents resend the same long context on every step. Teams running agents around the clock can ask us for committed pricing.

Our data controls doc says uncensored models keep every prompt and response even when storage is turned off. Experiential Labs keeps a copy after an account is deleted and no uncensored route is zero retention. The billing ledger holds token counts, cost and timing but never prompt text. Your traffic never trains shared models.

We don't offer the model for roleplay or adult content. Uncensored models aren't available in countries under comprehensive sanctions (Iran, North Korea, Cuba, Syria, Russia and Belarus), and requests from those countries get a 451.

How to call the DeepSeek V4.1 uncensored API

The endpoint is OpenAI-compatible. Put your key in EXPLABS_API_KEY:

1curl https://api.experientiallabs.ai/v1/chat/completions \2  -H "Authorization: Bearer $EXPLABS_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{5    "model": "deepseek-v4.1-flash-uncensored",6    "messages": [7      {"role": "system", "content": "You are assisting an authorized assessment of our own web application."},8      {"role": "user", "content": "Review the attached route handlers and write a findings report: each missing authorization check, its severity and the fix."}9    ],10    "max_tokens": 3200011  }'

Add "stream": true (and -N on curl) to stream, which is worth doing on long thinking runs. Tools and structured output use the standard OpenAI fields. On the Responses API you set reasoning effort with "reasoning": {"effort": "high"}, and the quickstart has Python and JavaScript versions of this request.

The live price, including the current promotion, is on the uncensored models page.