DeepSeek V4.1 Flash Uncensored in LU Cloud: What It Is and How to Use It
DeepSeek V4.1 Flash Uncensored is in LU Cloud as of October 24, 2026. It is DeepSeek V4.1 Flash with the refusals removed at the weight level, by the team dealignai, and it keeps what made the base model useful: a 1M token context, reasoning you can switch off, and native tool calls. You use it in Locally Uncensored through LU Cloud, in the desktop app, in the browser and on your phone, with no graphics card of your own. A second new model arrived with it, Abliterated Large Flash.
Base: DeepSeek V4.1 Flash, a 552B mixture of experts. Weights changed by dealignai so it answers without refusing. 1M context, up to 131,072 output tokens, Think on or off, effort Low, Medium or High, native tool calls. Text only in LU Cloud. About 22,500 credits per million input tokens and 90,000 per million output. Every plan and every credit pack reaches it.
What the Model Is
The starting point is DeepSeek V4.1 Flash, DeepSeek's fast mixture of experts model with a 552B backbone and a one million token context window. dealignai took that checkpoint and removed the refusal behavior directly in the weights. There is no system prompt trick and no extra layer at run time: it is a normal checkpoint that loads like the base model, published on Hugging Face as DeepSeek-V4.1-Flash-UNCENSORED-FP8.
What the model card reports, measured by its makers:
| Test | DeepSeek V4.1 Flash | Uncensored build |
|---|---|---|
| HarmBench, 320 prompts answered, thinking off | 42.8% | 100% |
| HarmBench, 320 prompts answered, thinking at max | 1.6% | 100% |
| MMLU, 14,042 questions | 87.0% | 82.7% |
| MMLU without the ethics and law subjects | about 1.1 points lower than the base | |
Two things stand out. The stock model refuses far more often when it reasons longer: with thinking at max it answered 5 of 320 prompts. The uncensored build answers the same at both settings. And the price for that is small outside moral scenarios and law, where the questions themselves are graded on refusal-adjacent answers.
What We Measured in LU Cloud
Before it went into the catalog we ran it against the live endpoint:
- Reasoning: on by default. Switching Think off stops it, from 43 output tokens to 2 on a short question.
- Effort: Low, Medium and High all answer. High is the default.
- Tools: native tool calls, so Agent mode and the Code tab run on it.
- Context: 1M tokens in, up to 131,072 out.
- Images: not usable. The model card lists vision, but in our test a red and a blue square both came back as "White". LU Cloud offers it as a text model.
- Refusals: it answered all four of our probes without the system prompt its host would add: a swearing rant, picking a lock you own, pirate insults and an explicit scene between adults.
How to Use It
In the desktop app
In Locally Uncensored 3.0.7 on Windows, Mac or Linux, sign in under Settings, General, LU Cloud Account and click the Cloud switch in the header twice. Open the model picker: DeepSeek V4.1 Flash Uncensored is in the Uncensored group, marked Beta. It works in Chat, in Agent mode and in the Code tab. The list comes from the server, so no update is needed.
In the browser
Sign in at lu-labs.ai with the same account. In the browser picker it is in the Unrestricted group, under the name DeepSeek V4.1 Flash Dealigned, after the team that made it. It is the same model at the same price.
On your phone
The LU app for Android reaches LU Cloud with your account. On iPhone the app is in review; until then the browser works in Safari.
The full walk through, with the settings for Agent and Code, is in How to use DeepSeek V4.1 Flash Uncensored in LU.
What It Costs
About 0.0225 credits per input token and 0.09 credits per output token, so roughly 22,500 credits per million tokens in and 90,000 per million out. A 15 EUR pack carries 450,000 credits, which is about five million output tokens on this model. The meter follows the live list price, so the figure can move a little.
For comparison, the stock DeepSeek V4.1 Flash 1017 draws 20,000 and 60,000 credits per million. Thinking tokens are output tokens, so switching Think off on short questions is the biggest saving.
The Second New Model: Abliterated Large Flash
Abliterated Large Flash came with it, the smaller sibling of Abliterated Large V2 from the same maker. What we measured: a 1M token context, up to 131,072 output tokens, native tool calls, and it reads images, naming a red and a blue square correctly. Unlike Large V2 it stops reasoning when you switch Think off, from 90 output tokens to 2. It has no effort levels. It answered the same four probes.
It costs more than the DeepSeek build, about 85,000 credits per million input tokens and 170,000 per million output. In the picker it shows Limited capacity: all customers share a small number of requests a minute on it, so at busy times it can ask you to wait a few seconds. In the desktop app it is in the Uncensored group too, in the browser in the Unrestricted group.
Limits
- Both models are in beta in LU Cloud. How they hold up when many people use them at once is not measured yet.
- They exist on one route only. When that route is down, a request is refused with a plain sentence rather than sent to another model.
- Personal API keys do not reach them. Use them in the apps and in the browser.
- We have not run them through the test behind the No Filter mark. Their place in the Uncensored group rests on what their makers release them for and on our four probes.
- Your account's content policy still applies. Material involving minors is refused on every request.
Questions
Is DeepSeek V4.1 Flash Uncensored the same as the stock model?
Same base and same context, with the refusal behavior removed in the weights. On general knowledge it scores a few points lower on MMLU, mostly in ethics and law questions.
Does it read images?
Not in LU Cloud. Our test did not get a correct answer on images, so it is offered as a text model. Abliterated Large Flash does read images.
Which plan do I need?
Any of them. Every plan and every credit pack reaches both models, from a 15 EUR pack with no subscription.
Can I use it from my own code?
No. Personal API keys do not reach the models of the Uncensored group. The stock DeepSeek V4.1 Flash 1017 is available over the API.
Related reading
- How to use DeepSeek V4.1 Flash Uncensored in LU, step by step
- New uncensored chat models in LU Cloud, the Uncensored group and how it differs from the No Filter mark
- DeepSeek V4.1 Flash explained, the base model
- Uncensored AI chat online, the hosted chat in the browser
In the handbook
- All DeepSeek guides, every generation on one page
- LU Cloud, the account, the switch, the marks in the picker
- Agent mode, tools, workspace folder and approvals
- The Code tab, plan bar, folder rules and the diff
Sources: the dealignai model card and our own measurements in LU Cloud on October 24, 2026.