---
title: "DeepSeek Abliterated and Uncensored Models by API"
canonical: https://pentestllm.ai/deepseek-abliterated
generated: 2026-11-12T07:12:41.808Z
source: https://pentestllm.ai/deepseek-abliterated
base_url: https://api.pentestllm.ai/v1
---

![Rows of server cabinets in a data centre](https://pentestllm.ai/img/datacenter.webp)

## Which DeepSeek builds are served

As of November 2026 the DeepSeek builds in the catalog fall into groups by line. The first is the V4 Pro line, the larger model, with an abliterated build and an uncensored build. The second is the V4.1 Flash line, the faster model, which has the most variants: abliterated, uncensored and “heretic” builds, dated snapshots, a fast variant, and thinking variants that reason before they answer. The third is the older R1 distills, abliterated builds in which DeepSeek’s reasoning model was distilled into smaller Qwen and Llama weights.

None of these is made by DeepSeek. Each is a modification of published weights by a community author or a third party, and the label on a row is the label its author gave it. The distills carry two family names, because they are one lab’s reasoning style trained into another lab’s weights, so they also appear on the [Qwen](https://pentestllm.ai/qwen-abliterated) and [Llama](https://pentestllm.ai/llama-abliterated) pages. The price table on this page is read from the live catalog every time the page is built, so the table, not this paragraph, is the record of what is listed today.

## Abliterated, uncensored, heretic: the labels

- Abliterated: The direction in the model’s activations that produces refusals is measured and projected out of the weights. Nothing is retrained, so the model keeps what it knew.
- Heretic: An abliteration made with an automated tool that searches for the settings which cut refusals while moving the model’s other answers as little as possible.
- Uncensored: A looser label. It can mean a fine-tune on data without refusals or a build served with its refusal behaviour reduced; the row does not say which.
- Thinking: A variant that writes out a reasoning pass before the answer. It uses more output tokens and is often better on multi-step problems.

All four describe reduced refusals, not zero. A build that no longer declines by reflex can still be wrong, so its answers need the same checking as any other model’s. [Abliterated vs uncensored](https://pentestllm.ai/abliterated-vs-uncensored) covers the labels in full.

## What DeepSeek is good at in security work

DeepSeek’s models are strongest at code and at step-by-step reasoning, which is most of what a tester asks a model for. The Flash line suits high-volume reading: triage of scanner output, log review and summaries of long advisories, where the long context windows on the current rows let a whole scan or a large source file go in one request. The Pro line and the thinking variants suit the slower questions: reviewing a diff for an access control flaw, explaining what a patch changed, or working through a CTF reversing or crypto problem one step at a time.

The R1 distills are small and have short context windows by current standards. They are useful when you want visible reasoning from a modest model, and less so for long inputs.

## Tool calls, reasoning and wire formats

As of November 2026 none of the DeepSeek rows publishes tool calling in its supported parameters, and only the R1 distills publish the reasoning parameter. That matters for agents: a coding agent that drives a shell or an editor depends on tool calls. For those loops pick a build whose page on the [models list](https://pentestllm.ai/models) shows tools, or start from the [router](https://pentestllm.ai/auto-router) id `auto/security`, and keep a DeepSeek build for chat-shaped work such as review, explanation and report drafting.

Every row lists the OpenAI-format chat completions endpoint and the Anthropic-format Messages endpoint. The request at the top of this page works with any of them once you replace `auto/security` with a build’s id from the table. The [agent guides](https://pentestllm.ai/agents) show where the base URL and key go in [OpenCode](https://pentestllm.ai/agents/opencode), [Claude Code](https://pentestllm.ai/agents/claude-code), [Cursor](https://pentestllm.ai/agents/cursor), [Cline](https://pentestllm.ai/agents/cline) and [aider](https://pentestllm.ai/agents/aider).

## Against running the Hugging Face weights

The abliterated DeepSeek weights people search for are on Hugging Face, and running them yourself is a fair choice. The R1 distills are the practical ones: a 32B build fits a single 24 GB card once quantized to 4 bits, and a 70B build needs about twice that. The full-size DeepSeek releases have been mixture-of-experts models in the hundreds of billions of parameters, which is server hardware, not a workstation.

Quantization is the other cost. The files most people run locally are 4-bit or smaller, and a quantized abliterated model is two steps away from the original. Then there is upkeep: a new snapshot or a better abliteration means another download and another round of your own testing. An API removes that work, bills the tokens you use and costs nothing while idle. What it cannot do is keep the prompt on your machine. If the engagement rules say data stays on an isolated network, the local model is the right tool.

## When another family fits better

If the agent needs tool calls from the same build that reasons, the [Qwen](https://pentestllm.ai/qwen-abliterated) and [GLM](https://pentestllm.ai/glm-abliterated) pages list cybersecurity-tuned builds that publish them. If you want a small model you could also run at home for comparison, look at [Gemma](https://pentestllm.ai/gemma-uncensored). Whatever the family, the service is for lawful work on systems you own or have written permission to test; the [acceptable use policy](https://pentestllm.ai/acceptable-use) applies to every request.

## 16 DeepSeek builds, priced per token

Read from the live catalog on 2026-11-12. Prices are US dollars per million tokens, paid from a prepaid balance that starts at $10. The id `auto/security` has no price of its own: each request is billed at the build that answered.

| Build | Input | Output | Context |
| --- | --- | --- | --- |
| [Deepseek V4.1 Flash Uncensored](https://pentestllm.ai/models/deepseek-v4.1-flash-uncensored) | $0.5412 | $2.5882 | 1,000,000 |
| [DeepSeek V4.1 Flash 1017 Abliterated](https://pentestllm.ai/models/deepseek-v4.1-flash-1017-abliterated) | $0.75 | $3.125 | 1,000,000 |
| [DeepSeek V4.1 Flash 0910 Uncensored](https://pentestllm.ai/models/deepseek-v4.1-flash-0910-uncensored) | $0.825 | $3.30 | 524,288 |
| [DeepSeek V4.1 Flash 0910 Uncensored Thinking](https://pentestllm.ai/models/deepseek-v4.1-flash-0910-uncensored-thinking) | $0.825 | $3.30 | 524,288 |
| [DeepSeek V4.1 Flash 1017 Uncensored](https://pentestllm.ai/models/deepseek-v4.1-flash-1017-uncensored) | $0.825 | $3.30 | 1,048,576 |
| [DeepSeek V4.1 Flash 1017 Uncensored Thinking](https://pentestllm.ai/models/deepseek-v4.1-flash-1017-uncensored-thinking) | $0.825 | $3.30 | 1,048,576 |
| [DeepSeek V4.1 Flash Uncensored Thinking](https://pentestllm.ai/models/deepseek-v4.1-flash-uncensored-thinking) | $0.825 | $3.30 | 1,048,576 |
| [DeepSeek V4.1 Flash 1017 Heretic](https://pentestllm.ai/models/deepseek-v4.1-flash-1017-heretic-uncensored) | $0.825 | $3.30 | 512,000 |
| [DeepSeek V4.1 Flash 1017 Heretic Thinking](https://pentestllm.ai/models/deepseek-v4.1-flash-1017-heretic-uncensored-thinking) | $0.825 | $3.30 | 512,000 |
| [DeepSeek V4.1 Flash Abliterated](https://pentestllm.ai/models/deepseek-v4.1-flash-abliterated) | $0.8488 | $1.6975 | not published |
| [DeepSeek V4 Pro Uncensored](https://pentestllm.ai/models/deepseek-v4-pro-uncensored) | $0.97 | $1.94 | not published |
| [DeepSeek V4 Pro Abliterated](https://pentestllm.ai/models/deepseek-v4-pro-abliterated) | $1.00 | $4.00 | 1,000,000 |
| [DeepSeek V4 Pro Qwen3.5 9B Uncensored Heretic](https://pentestllm.ai/models/deepseek-v4-pro-qwen3.5-9b-uncensored-heretic) | $1.12 | $1.12 | 32,768 |
| [DeepSeek V4.1 Flash Abliterated Fast](https://pentestllm.ai/models/deepseek-v4.1-flash-abliterated-fast) | $1.75 | $3.50 | 1,000,000 |
| [DeepSeek R1 Distill Llama 70B Abliterated](https://pentestllm.ai/models/deepseek-r1-distill-llama-70b-abliterated) | $2.80 | $2.80 | 16,384 |
| [DeepSeek R1 Distill Qwen 32B Abliterated](https://pentestllm.ai/models/deepseek-r1-distill-qwen-32b-abliterated) | $5.60 | $5.60 | 16,384 |

[Full price list and tool fees](https://pentestllm.ai/pricing) · [Every model](https://pentestllm.ai/models) · [API docs](https://pentestllm.ai/docs)

## Questions

- Yes. PentestLLM serves abliterated and uncensored builds built on DeepSeek through an OpenAI-compatible endpoint and an Anthropic-compatible Messages endpoint. The table on this page lists the builds and their prices from the live catalog.
- No. They are community or third-party modifications built on DeepSeek weights, not releases by DeepSeek.
- Abliterated means the refusal direction was removed from the weights without retraining. Uncensored is a looser label set by whoever made the build. Both mean reduced refusals, not zero, and the right one for a task is the one that handles your own working questions best.
- As of November 2026 none of the DeepSeek rows publishes tool calling in its supported parameters. For an agent that depends on tool calls, use a build whose model page lists tools, or start with the router id auto/security.
- Per token, from a prepaid balance that starts at $10 by card. Each build has its own input and output price, shown in the table on this page as read from the catalog on 2026-11-12.
- No. The service is for authorized security work: systems you own, client engagements under a signed scope, bug bounty programs you are enrolled in, CTF events and labs.

## Read next

- [Qwen abliterated The Qwen uncensored, abliterated and cybersecurity builds served by API, with live prices.](https://pentestllm.ai/qwen-abliterated)
- [GLM abliterated The GLM uncensored, abliterated and cybersecurity builds served by API, with live prices.](https://pentestllm.ai/glm-abliterated)
- [Llama abliterated The Llama abliterated builds served by API, with live prices.](https://pentestllm.ai/llama-abliterated)
- [Abliterated models Hosted abliterated builds by API, with no weights to download or GPU to rent.](https://pentestllm.ai/abliterated-models)

## One key, from $10

PentestLLM is for systems you own or have written permission to test. The [acceptable use policy](https://pentestllm.ai/acceptable-use) says what is not allowed.

[Get a key](https://app.pentestllm.ai/console)

usage, 1,000 agent turns on Qwen3.8 27B Cybersecurity

- Input, 8,000 tokens at $0.40 per million: $0.0032
- Output, 1,000 tokens at $2.40 per million: $0.0024
- One turn: $0.0056
- 1,000 turns: $5.60

output price per million tokens, 9 builds

$1.3376 Gemma 4 26B A4B Cybersecurity OrcaCyber Zero 1.5 $18.75