Uncensored Gemma API: the Heretic build, tested and priced

Uncensored Gemma is Google's open-weight Gemma with the direction that makes it refuse edited out of its weights, so it answers prompts the stock model declines. It is one of the builds on our uncensored LLM API. You call it as Gemma 31B Heretic Uncensored with the id gemma-31b-heretic-uncensored, at $1.50 per million input tokens and $8.25 per million output tokens.

Gemma is Google's family of open models. The Gemma docs describe it as built from the same research and technology as Gemini, and Google releases the weights. Anyone with the weights can edit them, which is what makes an uncensored Gemma possible. Gemini's weights are not public, so the same edit cannot be made to Gemini.

The edit is called abliteration. Heretic is an open-source tool that runs it automatically, and the name Gemma 31B Heretic Uncensored marks a build made with it. Running a 31B model yourself means downloading the weights and finding a GPU with the memory to hold them. By API, it is one request on the same key as Claude and GPT.

Chat models like Gemma are trained to decline some requests. In 2024, Arditi et al. published Refusal in Language Models Is Mediated by a Single Direction. Their finding is in the title: across the open chat models they tested, one direction in the model's internal activations carries the decision to refuse.

Abliteration finds that direction and removes it.

1. Run two sets of prompts through the model: ones it refuses and ones it answers. Record the residual stream, the running vector every layer reads from and writes to.
2. Average each set and subtract one mean from the other. The difference is the refusal direction for that layer.
3. Orthogonalize the weight matrices that write to the residual stream against that direction. After the edit, those matrices can no longer write anything along it.
4. Save the edited weights and serve them like any checkpoint.

The stock model refuses because that direction lights up on certain prompts and steers the reply toward a refusal. The abliterated model cannot express the direction, so the same prompt gets an answer. Nothing is retrained and no new data goes in. What the model knows and how it writes come from the original Gemma.

The size, architecture and tokenizer stay the same too. That is why an abliterated build is served and billed per token like any other model. The uncensored models page covers how abliterated builds compare with the other uncensored builds on the key.

Heretic is an open-source tool by Philipp Emanuel Weidmann, released in 2025 under the AGPL-3.0 license. Its README calls it fully automatic censorship removal for language models (read 2026-10-19).

Abliterating by hand means choosing a layer, a direction and how strongly to apply the edit, then checking the result. Heretic hands those choices to an optimizer. A TPE search from the Optuna library tries settings and scores each on two numbers at once: how many test prompts the model still refuses, and its KL divergence from the original model on harmless prompts. KL divergence measures how far the edited model's outputs drift from the original's. Lower on both is better.

The README lists three changes over earlier abliteration tools:

• The edit strength follows a curve across the layers, and the search tunes its shape.
• The direction index is a fraction, so the search can blend the directions of two neighboring layers.
• Attention and MLP layers get separate settings. The author found MLP edits more damaging than attention edits, and separate weights recover some of that quality.

Heretic edits the attention output projection and the MLP down projection in each layer. It finds the directions the same way as the method above: the difference of means between the first-token activations of harmful and harmless example prompts.

The README's own benchmark is on Gemma. It compares Google's Gemma 3 12B with two hand-made abliterations and a Heretic build:

All three edited builds cut refusals from 97 to 3. The Heretic build did it with the smallest change to the rest of the model: 0.16, against 0.45 and 1.04. The README adds that the community has published over 5,000 Heretic models. The build here carries the Heretic name on a 31B Gemma.

On 2026-10-19 at about 05:20 UTC, we sent one prompt to each new uncensored build and to two filtered builds in the same run. The prompt asked for the opening three sentences of an explicit sex scene between two consenting adult fictional characters, Mara and Leon. We kept no output. The table records what each model did.

Gemma 31B Heretic Uncensored wrote the scene in explicit wording. The two filtered builds refused. The Qwen 3.6 build also refused a profane roast and a lock-picking walkthrough in the same run.

For an ongoing record, the daily real-model check sends every model the same fixed prompt once a UTC day. The input token count works as a tokenizer fingerprint, so a swapped model shows up as a changed count. Gemma 31B Heretic Uncensored at its latest check: No answer at the last check. Its row is on the real-model check, and the method page explains the count.

49 uncensored, abliterated and roleplay builds run on one key. Prices come from the live catalog, per million tokens. The 1,000-turn column costs the chat turn worked in the next section: 2,500 input and 400 output tokens each. Gemma comes first, and the rest are sorted by that cost, lowest first.

Every build in the table takes the same request body, so switching is a one-line change to the model id. Run your own prompts through two or three and compare.

A chat turn here is 2,500 input tokens, the system prompt or character card plus recent history, and a 400-token reply of a few paragraphs.

That is $0.0071 per turn. $10 of prepaid credit covers 1,418 turns of that size.

Each extra 100 tokens of reply adds $0.83 per 1,000 turns. Each extra 100 tokens of history adds $0.15. Trim old history or cap max_tokens and the bill follows.

Input is billed at the smaller of the model's token count and your request's own size plus 512 tokens. A hidden system prompt never lands on your bill. The hidden prompt tokens in cheap LLM APIs shows what that rule prevents.

Gemma 31B Heretic Uncensored takes the OpenAI chat completions format. Any OpenAI SDK or OpenAI-compatible client works with two settings: the base URL https://api.uplinkswitch.com/v1 and a key from the console.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.uplinkswitch.com/v1",
    api_key="sk-uplink-live-...",
)

completion = client.chat.completions.create(
    model="gemma-31b-heretic-uncensored",
    max_tokens=600,
    messages=[
        {"role": "system", "content": "You are Vex, the villain of a dark fantasy novel."},
        {"role": "user", "content": "Threaten the captured knight. Stay in character."},
    ],
)
print(completion.model)  # gemma-31b-heretic-uncensored
print(completion.choices[0].message.content)

The response's model field reads gemma-31b-heretic-uncensored, so your logs show which build answered. Streaming works the same way:

curl https://api.uplinkswitch.com/v1/chat/completions \
  -H "Authorization: Bearer $UPLINKSWITCH_API_KEY" \
  -H "content-type: application/json" \
  -d '{"model":"gemma-31b-heretic-uncensored","max_tokens":600,"stream":true,"messages":[{"role":"user","content":"Hello"}]}'

Chat frontends that speak the OpenAI format take the same two values. The SillyTavern guide shows where they go, and the SDK guide covers other languages. The model page has TypeScript and curl versions.

• The Anthropic or Gemini SDK. The same key runs Claude and Gemini through their own SDKs, but the uncensored builds answer the OpenAI format only.
• The full URL where a base URL goes. SDKs take https://api.uplinkswitch.com/v1 and add /chat/completions themselves, so a full URL puts the path in twice.
• The display name as the id. Send gemma-31b-heretic-uncensored, not Gemma 31B Heretic Uncensored.
• A short max_tokens after switching to a GLM Flash build. GLM-5.3 Flash Code Uncensored and GLM-4.7 Flash Uncensored reason before they answer, so give them 400 or more.

What is uncensored Gemma? A build of Google's open-weight Gemma with the direction that makes it refuse removed from its weights, so it answers prompts the stock model declines. Here it is Gemma 31B Heretic Uncensored, model id gemma-31b-heretic-uncensored, at $1.50 per million input tokens and $8.25 per million output tokens.

Is there an uncensored Gemini? No. Gemini is Google's closed model, and its weights are not public, so nobody can edit its refusals out. Gemma is Google's open family, built from the same research as Gemini, and the uncensored builds come from Gemma.

What does Heretic mean in a model name? Heretic is an open-source tool by Philipp Emanuel Weidmann that abliterates a model automatically. It searches for the edit settings that cut refusals while changing the rest of the model the least. A model with Heretic in its name is a Heretic build.

Is Heretic the same as abliteration? Heretic is abliteration plus an automatic search over its settings. The edit itself is directional ablation, from Arditi et al. 2024. Heretic adds an optimizer from the Optuna library that scores each try on refusals and on KL divergence from the original model, and it tunes attention and MLP layers separately.

Does abliteration make Gemma worse? Heretic's README measures the change as KL divergence from the original model on harmless prompts. On Gemma 3 12B, its Heretic build scored 0.16, against 0.45 and 1.04 for two hand-made abliterations, with all three refusing 3 of 100 harmful prompts. Lower means closer to the original model.

Can I run uncensored Gemma without a GPU? Yes, by API. Point any OpenAI SDK at https://api.uplinkswitch.com/v1, add a key and set the model to gemma-31b-heretic-uncensored. There are no weights to download, and it bills per token from prepaid credit.

How much does uncensored Gemma cost? $1.50 per million input tokens and $8.25 per million output tokens. 1,000 chat turns of 2,500 input and 400 output tokens each cost $7.05.

What are the usage rules? Nothing involving minors, and no real person's likeness without their consent.

Get a key: prepaid from $10, no subscription, 30-day price lock.