Gemma Uncensored — Google's Gemma 4, Without the Refusals

Gemma is Google's open-weights family — capable, efficient, and famously heavily aligned: the stock instruct models refuse a wide band of legal prompts. Uncensored community builds fix that. This page covers the hosted Gemma 4 uncensored build you can use right now, and how it compares to the stock model.

Chat anonymously…

100% anonymous·Zero logs·Encrypted

You can chat with an uncensored build of Google's Gemma 4 online at Pinkerton AI: a 26B mixture-of-experts model (25.2B total, 3.8B active parameters) with refusal behaviour removed, served end-to-end encrypted into a Trusted Execution Environment with hardware attestation. No account is required, no server-side logs are kept, and it costs 3 credits per message with a free tier.

Last updated 2026-08-17

request_flow.diagram3 steps

REQUEST

Your prompt

Any topic, any tone — no pre-filter.

PIPELINE

No moderation layer

Skips the refusal/disclaimer pass mainstream models run.

RESPONSE

Direct, complete answer

Nothing held back, nothing logged.

What is Gemma uncensored?

A community-modified build of Google's Gemma 4 in which the trained refusal behaviour has been removed, while the model's knowledge and writing quality stay intact. Stock Gemma is one of the most conservative open models — it refuses or moralises on dark fiction, blunt opinions, security detail and most mature themes. The uncensored build answers those directly, within the law. The variant hosted here is the 26B mixture-of-experts Gemma 4: 25.2B total parameters with 3.8B active per token, so it runs fast for its quality class.

What makes this hosting unusual: E2EE into a TEE

The Gemma uncensored endpoint runs inside a Trusted Execution Environment — an encrypted hardware enclave whose integrity can be verified through attestation. Your messages are encrypted end-to-end into the enclave, which means even the server operator cannot read the exchange in transit or in memory. Combined with Pinkerton AI's no-account, no-logs design, this is about as private as hosted inference gets: nothing readable server-side, nothing stored, nobody to attach it to.

How does it compare to stock Gemma 4?

Identical base capability — same architecture, same knowledge. The difference is behavioural: where stock Gemma inserts refusals, safety disclaimers and tone-policing on legal prompts, the uncensored build simply answers. If you've used Gemma locally and liked its speed and writing but hit its refusal wall, this is the same model with the wall removed. Illegal content remains refused; uncensoring removes editorial judgment, not the law.

Gemma uncensored online vs running it yourself

Uncensored Gemma builds circulate on Hugging Face and run well locally — the MoE design is efficient enough for a mid-range GPU at Q4. Local wins on absolute privacy; hosted wins on setup (none), speed (datacenter hardware), and here specifically on the E2EE/TEE guarantee, which most self-hosted setups don't replicate for remote access. If you just want to talk to Gemma without refusals in the next thirty seconds, hosted is the practical answer.

Getting started

Open the chat, pick 'Gemma 4 26B A4B Uncensored' from the model picker, and start — no account, no email. It costs 3 credits per message; a free session includes credits to test it against the default GLM Heretic and Venice Uncensored 1.2. If the 64K E2EE context is limiting for your use case, the other uncensored models here offer up to 200K without the enclave.

Pinkerton AI vs Stock Gemma 4: how do they compare?

Feature	Pinkerton AI	Stock Gemma 4
Refusal behaviour	Removed — answers legal prompts directly	Frequent refusals and hedging
Architecture	Gemma 4 MoE, 25.2B total / 3.8B active	Same base architecture
Hosting	E2EE into a TEE, hardware attestation	Wherever you deploy it
Setup	None — select and chat, no account	Self-host or cloud deploy
Context window	64K on the E2EE endpoint	Up to 256K self-hosted
Cost	3 credits/message, free tier	Free weights + your hardware

FAQ

Is there an uncensored version of Google's Gemma?

Yes — community builds of Gemma 4 with refusal behaviour removed. Pinkerton AI hosts the 26B MoE uncensored build, chattable online with no account.

Is Gemma uncensored free to use?

A free session with credits and no sign-up lets you test it; it costs 3 credits per message afterwards, with plans from €17.99/month.

What does the E2EE/TEE hosting actually guarantee?

Messages are encrypted end-to-end into a Trusted Execution Environment with hardware attestation — the exchange is unreadable server-side, even by the operator, and no conversation log is kept.

Is uncensored Gemma as smart as stock Gemma?

Yes — the modification removes refusal behaviour, not capability. Knowledge and writing quality are those of Gemma 4 26B MoE.

Can I run Gemma uncensored locally instead?

Yes — uncensored Gemma builds are on Hugging Face and run under Ollama or llama.cpp on a mid-range GPU. Hosted access here is for when you want it instantly, with E2EE, and at full serving speed.

Start chatting — free to try

No signup. No email. Just open the chat.