Gemma uncensored, hosted by API

Gemma attracts a great many community tunes, and most of the ones in circulation are tuned for fiction writing. This site does not list those. What it lists are the Gemma builds meant for work: as of November 2026 that is a cybersecurity-tuned build on Gemma 4, abliterated builds on the larger Gemma 4 model including a heretic build, and uncensored tunes on Gemma 4 and Gemma 3.

All of them are built on Gemma by community authors; Google publishes the base models, not these. The table on this page is read from the live catalog and shows the builds that match today.

Gemma is a compact family. The builds here are in the 26B to 31B class, which makes them quick. That suits the repetitive parts of an engagement: classifying findings, turning tool output into a ranked list, rewriting notes into report language, and answering short questions about a CVE advisory. Some of the Gemma 4 builds also list image input, which helps when the evidence is a screenshot.

It is less suited to the hardest reasoning. For a hard reversing problem, a multi-stage CTF challenge or a review of a large codebase, a bigger model will usually hold the thread better. A common pattern is a small build for volume and a larger build for the few requests that need depth, which is what the router id auto/security does across the security builds.

Uncensored

A fine-tune that reduces refusals, usually by training on data in which the model answers. The label is the author’s, and methods vary.

Abliterated

The refusal direction is removed from the weights with no retraining. Heretic builds do this with an automated search for the least disruptive settings.

Cybersecurity

A fine-tune on security material. It answers in the field’s terms, knows the common tooling, and declines security questions less often.

For pentest work the cybersecurity build is the default. Stock Gemma has a reputation for caution, so the difference between the stock model and these builds is easy to notice on security questions. It is still reduced refusals, not zero.

As of November 2026 the cybersecurity-tuned build publishes tool calling and a reasoning parameter. Among the uncensored tunes, one publishes tool calling with structured output against a JSON schema and another publishes tool calling with reasoning. The abliterated builds publish no tool calling, though one publishes a reasoning parameter, so keep them for chat-shaped work.

Every row lists the OpenAI-format chat completions endpoint and the Anthropic-format Messages endpoint. The curl request at the top of this page is the OpenAI shape; replace auto/security with a Gemma id from the table to pin a build. Setup for OpenCode, Claude Code, Cursor, Cline and aider is in the agent guides.

Of the families on this site, Gemma is the one you can most easily run yourself. A build of this size fits a single 24 GB card at 4-bit quantization, the uncensored and abliterated weights are on Hugging Face in ready-made files, and for work that must stay offline that is the right setup.

The API is worth it in narrower cases than for the giant models. You get the full context window without trading it against GPU memory. You can try several tunes in an afternoon without downloading each one. A laptop on a client site can use the model with no workstation behind it. You pay for the tokens used and nothing when idle. Prompts do leave your machine, so check your engagement rules first.

For long inputs and agent loops, the Qwen and GLM pages list cybersecurity builds on larger models. For the classic abliterated 70B models, see Llama. All of it is sold for lawful work on systems you own or have written permission to test, under the acceptable use policy.