Google's Gemma 4 ships safety-tuned and refuses adult roleplay. The community fine-tunes built on it do not, and the best of them are the most natural persona chat models available at their size. We explain what a fine-tune actually changes, why the 26B class is the sweet spot, how "for her" and "for him" tunes differ, and the failure modes to watch.
Short answer: base Gemma 4 refuses adult content; the open weights make that fixable, and the roleplay fine-tunes built on it are among the best persona chat models you can run yourself. They keep Gemma’s fluency and emotional range, drop the refusals, and add the pacing of actual roleplay. MyBabes runs several 26B-class Gemma 4 tunes for persona chat, including separate tunes for female and male personas and a mixture-of-experts variant, alongside a Qwen 3.5 writer tune for long-form scenes.
Why Gemma at all
Google’s Gemma family is unusual among open-weights LLMs for how natural its conversational register is: it handles tone, emotional subtext and short-message rhythm better than most models of similar size, which are tuned for assistant tasks. That is exactly what persona chat needs. The base model’s safety tuning makes it refuse anything sexual, and its chat template is strict about turn order, but both are solvable, and the community solved them.
What a fine-tune changes, and what it does not
An uncensored fine-tune retrains the instruction layer on roleplay and adult conversation data. It changes:
Refusals. Removed, along with the moralising preambles.
Register. The model learns to stay in character, escalate at the pace of the conversation and use explicit language when the scene calls for it.
Length discipline. Good tunes learn to vary reply length; base models drift toward essays.
It does not change:
Knowledge and reasoning. Same base capability.
Context handling. Same context window and the same tendency to lose track of details deep into a long chat.
The template. Gemma’s alternating-turn requirement remains; the serving stack has to handle it.
Why 26B is the sweet spot
Size class | Persona quality | Cost to serve | Notes |
|---|---|---|---|
4–9B | Serviceable, repetitive | Very low | Forgets details, loops phrases |
12B | Good | Low | The budget tier for many platforms |
26B (dense or MoE) | Very good | Moderate | Best balance; MoE variants serve cheaper |
70B+ | Excellent | High | Diminishing returns for chat length replies |
The 26B class is where the tunes stop sounding like a model and start sounding like a person, while still serving at a cost that works for a consumer product with thousands of concurrent chats. The mixture-of-experts variants in this class (Serenity is the one we run) activate a fraction of the parameters per token, which keeps latency and cost closer to a 12B model.
Gendered persona tunes
The persona a user talks to has a gender, and the training data that produces a convincing female persona is not the same as for a male one: vocabulary, escalation patterns and what “in character” means differ. Some tuners now train separate variants. On MyBabes, the female personas run on a “for her” tune and the male personas on a “for him” tune, both Gemma 4 26B class. In blind comparison against a single general tune, the gendered variants scored higher on staying in character and on natural escalation, and lower on the generic “AI-ness” that breaks immersion.
What we measured
Persona chat evaluation on a fixed set of 40 scenarios, three tunes, scored by reviewers on immersion, escalation pacing, consistency and explicitness when warranted.
General Gemma 4 tune | Gendered Gemma 4 tunes | Qwen 3.5 27B Writer | |
|---|---|---|---|
Stays in character | 4.0 | 4.5 | 4.3 |
Escalation pacing | 3.8 | 4.4 | 4.1 |
Consistency over 50 turns | 3.6 | 3.9 | 4.2 |
Explicit when warranted | 4.2 | 4.6 | 4.4 |
Prose quality in long scenes | 3.7 | 3.9 | 4.7 |
Short-message rhythm | 4.5 | 4.6 | 3.9 |
The pattern: Gemma tunes win at chat, the Qwen writer tune wins at prose. That is why both are offered; our Qwen 3.5 analysis covers the other side.
Where the models break
Long-context drift. Past roughly 50 turns, details from early in the conversation get lost or contradicted. Summarised memory in the serving layer helps; the model alone does not solve it.
Repetition loops. Under aggressive sampling, tunes fall into repeated phrases. Repetition penalties and a capped reply length are mandatory.
Over-escalation. Some tunes go explicit too fast. The gendered tunes were better here, and the system prompt has to do the rest.
Template strictness. Gemma templates require the first turn to be the user’s and strict alternation; a greeting-first conversation needs handling in the serving stack or the model errors out.
Running one yourself
The tunes are on the usual model hubs in GGUF and safetensors form. A 26B dense model at Q8 needs about 28 GB of VRAM; Q4 fits in 16 GB with quality loss; the MoE variants need the full weights in memory but run faster. Serve with a Gemma-aware chat template and a system prompt that defines the persona, the register and the boundaries. Or use a platform that has done that work; persona chat is the core of MyBabes and the tunes above are the ones behind it.
Related in the Lab
Key takeaways
Base Gemma 4 refuses adult roleplay; the fine-tunes remove refusals and add roleplay pacing.
26B is the sweet spot for quality per serving cost; MoE variants make it cheaper.
Gendered persona tunes beat a general tune on immersion and escalation.
Gemma tunes win at chat rhythm; a Qwen writer tune wins at long prose.
Long-context drift and repetition need handling in the serving layer.
FAQ
Is Gemma 4 uncensored?
Not as released by Google. Community fine-tunes on the open weights remove the refusals and are used for adult roleplay.
Which Gemma 4 fine-tune is best for roleplay?
The 26B-class tunes, including gendered persona variants and the Serenity MoE tune, scored highest in our evaluation. Names change often; check the model hub for current versions.
Can I run an uncensored Gemma 4 locally?
Yes. A 26B tune needs roughly 16–28 GB of VRAM depending on quantisation.
Is Gemma 4 better than Qwen for NSFW chat?
For short-message persona chat, Gemma tunes rate higher. For long-form scenes and prose, the Qwen 3.5 writer tune wins.
Last tested: August 2026.
Skip the filter
Run these models with adult content enabled
MyBabes runs Wan 2.7, Seedance 2.0 and Krea inside its own generation stack, so the prompts refused on official apps generate as written. No GPU, no setup.
Open MyBabes
Builds MyBabes, runs the Lab
Jiri builds MyBabes and runs the Lab’s model testing: the same prompt set on every image, video and chat model, on the vendor’s official surface and on MyBabes, re-run after every release. He writes up what the models actually do, not what the launch posts say.
Follow on X
More from the Lab
Latest articles
All articles
Analysis
AI Porn and the Law in 2026: What's Actually Illegal, What's Regulated, and Where the Lines Sit
2025 and 2026 turned AI adult content from a legal grey zone into a regulated space with real enforcement: a US federal crime for non-consensual imagery, deepfake laws in 48 states, UK fines for missing age checks, EU labeling duties. Here is the map — and the one principle that organises all of it.
·
5 min read
How-To
Why AI Porn Looks Off: Every Artifact Explained, What Causes It, and How to Fix Each One
Everyone can tell when an AI image is off, few can say why. Each artifact has a specific technical cause and a specific fix — hands, limbs, faces, skin, teeth, text, seams, smear, flicker. This is the field guide, for images and video.
·
5 min read
Analysis
ChatGPT and NSFW in 2026: The Adult Mode That Never Shipped, What's Actually Allowed, and Where People Went Instead
"Treating adults like adults" was the promise in October 2025. Ten months later the adult mode is paused with no timeline, ChatGPT still refuses explicit content, and the search interest hasn't gone anywhere. Here is the full story, the current policy line in practice, and what the alternatives actually are.
·
4 min read