Mistral uncensored, by API
Mistral uncensored builds are third-party and community versions of Mistral’s models that decline fewer questions than the stock releases, and PentestLLM serves them by API so there are no weights to download. They are reached with one prepaid key from $10 through OpenAI-format and Anthropic-format endpoints, for authorized security testing only.
from openai import OpenAI
client = OpenAI(base_url="https://api.pentestllm.ai/v1", api_key="sk-ptl-live-…")
reply = client.chat.completions.create(
model="auto/security",
messages=[{"role": "user", "content": "Explain this advisory and list what to check on an affected host."}],
)
print(reply.model) # the build that answered
print(reply.choices[0].message.content)
The Mistral-family builds served
As of November 2026 the catalog carries uncensored builds on the Mistral Large line, on Magistral Small, which is the reasoning model, and on the small Ministral models, plus a Dolphin community tune built on a 24B Mistral model. All are labelled uncensored. None is tagged abliterated and none is a cybersecurity tune, so we do not describe them that way.
They are built on Mistral’s models by others, not released by Mistral. The price table on this page is filled from the live catalog: it lists the Mistral-family builds this site carries per token on the day, and shows the security builds when there are none.
How these builds are billed
One thing differs from the other families. As of November 2026 most of the Mistral-family rows are billed per request, not per token, and the table shows that flat fee in place of the token rates. Each build’s page under models states its unit and its figure, and every unit is paid from the same prepaid balance.
A flat fee per request suits long prompts, because the cost does not move with length. Per-token billing is better for many small requests. Check the unit before you wire a build into an agent that sends hundreds of calls in a session.
What the Mistral lines are good at
- Mistral Large
- The general flagship. Clear technical writing and strong multilingual output, which helps when a report or a client’s source material is not in English.
- Magistral
- The reasoning line. It works through a problem in steps before answering, which suits CTF challenges and questions about why a patch fixes a flaw.
- Ministral
- Small, fast models for bulk work: classifying findings, tagging log lines and first-pass summaries.
Mistral’s models are notably good in French, German, Spanish and other languages beside English. For a tester that is practical: findings drafted in the client’s language, and source comments, tickets and documentation in other languages read without a translation step.
Uncensored, not abliterated
An abliterated build has had the refusal direction removed from its weights. A security-tuned build has been trained on security material. Uncensored is looser than both: it says the build declines less and does not say how. The Dolphin tunes are a known example of one method, fine-tuning on data from which refusals were filtered out; for the other rows the catalog gives only the label.
So test before you rely on one. Put your ordinary working questions to it and see what it does. Expect reduced refusals, not zero, and if a build declines security material that matters to you, the security builds behind auto/security are tuned for it. Uncensored LLMs for pentesting compares the methods.
Tool calls and endpoints
As of November 2026 the Mistral-family rows publish no supported parameters: no tool calling, no reasoning parameter, no structured output. Use them for chat-shaped work such as explanation, review and drafting, and pick a build whose model page lists tools for an agent that needs them.
The rows list both the OpenAI-format chat completions endpoint and the Anthropic-format Messages endpoint. The Python request at the top of this page is the OpenAI shape against https://api.pentestllm.ai/v1. OpenCode, Cursor, Cline and aider use the same shape, and Claude Code uses the Messages endpoint.
Running Mistral weights locally
Mistral publishes open weights for much of its range, and the small end is easy to host. A Ministral model runs on a modest card or a recent laptop, and a 24B model such as Magistral Small fits a single 24 GB card when quantized. Uncensored community versions of those are on Hugging Face, and for offline work that is the right route.
The Large line is another matter: it needs server-class memory, and a heavily quantized copy gives up much of what makes it worth running. That is where an API is the practical choice. For the small models the API is a convenience: no download, no quantization decision, no runtime to keep patched, and a laptop on a client site can call it. With an API the prompt leaves your machine, so check your engagement rules.
When another family fits better
If you need an abliterated build specifically, see Llama or DeepSeek. If you need a cybersecurity-tuned build with tool calls, see Qwen or Gemma. Whatever you choose, the service is for lawful work on systems you own or have written permission to test, under the acceptable use policy.
6 Mistral 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 |
|---|---|---|---|
| Dolphin Mistral 24B Venice Edition | $0.50 | $2.25 | 32,768 |
| Magistral Small Uncensored | $0.005 a request | flat | not published |
| Ministral 3 14B Instruct Uncensored | $0.005 a request | flat | not published |
| Ministral 3 8B Instruct Uncensored | $0.005 a request | flat | 262,000 |
| Mistral Large 2 Uncensored | $0.005 a request | flat | 256,000 |
| Mistral Large 3 Uncensored | $0.09 a request | flat | 262,144 |
Questions
Yes. PentestLLM serves uncensored builds built on the Mistral Large, Magistral and Ministral lines through OpenAI-compatible and Anthropic-compatible endpoints, paid from a prepaid balance that starts at $10.
As of November 2026 the Mistral-family builds in the catalog are labelled uncensored, and none is tagged abliterated. The Llama, DeepSeek, Qwen and GLM pages list abliterated builds.
Not all of them. As of November 2026 most Mistral-family rows are billed per request and the rest per token. Each model page states its unit, and all of them are paid from the same prepaid balance.
No. They are third-party and community builds on Mistral models, not releases by Mistral.
As of November 2026 the Mistral-family rows publish no supported parameters, so do not rely on tool calling from them. For agent work use a build whose model page lists tools.
People doing authorized security work on systems they own or have written permission to test, including client engagements, bug bounty programs, CTF events and labs.
Read next
- Kimi uncensoredThe Kimi uncensored builds served by API, with live prices.
- Gemma uncensoredThe Gemma uncensored, abliterated and cybersecurity builds served by API, with live prices.
- Uncensored LLMThe uncensored and abliterated builds on one key, with live prices and context windows.
- Abliterated vs uncensoredThe difference between abliterated, uncensored, derestricted and security-tuned builds.
One key, from $10
PentestLLM is for systems you own or have written permission to test. The acceptable use policy says what is not allowed.
- 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