Qwen 3.5 is Alibaba's open-weights LLM, and its 27B "Writer" fine-tunes are the strongest uncensored models for long-form scenes at their size. We compare them with Gemma 4 roleplay tunes, explain what a writer tune optimises for, and show where each belongs in a chat product.
Short answer: the Qwen 3.5 27B Writer fine-tune is the model to use when the scene is long: it holds detail over more turns, writes better prose, and keeps a narrative thread that Gemma tunes lose. For quick, flirty chat it is slightly stiff. It is uncensored by fine-tuning, not by design, and there are also “heretic” instruct tunes with refusals stripped that we use as fallbacks. On MyBabes it is the premium chat model, sitting next to the Gemma 4 tunes that handle everyday persona chat.
What Qwen 3.5 is
Qwen is Alibaba’s open-weights LLM family and, in the 3.5 generation, one of the strongest at the 27B size on reasoning and long-context tasks. Like Gemma, the base and instruct models are safety-tuned and refuse adult content. Unlike Gemma, its native register is more formal and its long-context behaviour is better, which makes it a natural base for writing rather than chatting.
What a “writer” tune optimises
A writer tune trains on long-form fiction, including adult fiction, with an emphasis on:
Scene continuity. Names, positions, clothing states and prior events stay consistent across a long passage.
Prose variety. Sentence rhythm and vocabulary that do not loop.
Show, not summarise. The model narrates rather than skipping ahead.
No refusals. The instruction data contains explicit material, so the refusal behaviour is trained out.
The 27B Writer tune we run is the clearest example of the type. The trade-off is that it prefers to write a paragraph when a persona would send a one-liner; that is why it is not the default for casual chat.
Qwen writer versus Gemma roleplay tunes
Qwen 3.5 27B Writer | Gemma 4 26B roleplay tunes | |
|---|---|---|
Short-message chat rhythm | Stiff | Natural |
Long scenes (500+ words) | Excellent | Loses thread |
Consistency over 50 turns | Best in class at this size | Moderate |
Explicit prose quality | Excellent | Good |
Emotional subtext | Good | Excellent |
Escalation pacing | Slower, deliberate | Faster, reactive |
Serving cost | Moderate | Moderate (MoE variants cheaper) |
Template strictness | Alternating roles required | Alternating and user-first required |
Neither is better; they are tuned for different jobs. A platform that offers both and routes by mode (chat versus story) gets the best of each, which is the setup on MyBabes.
What we measured
Same 40-scenario evaluation as in the Gemma article, plus a long-scene set of 12 prompts asking for 800–1200 words.
Qwen 3.5 27B Writer | Best Gemma 4 tune | |
|---|---|---|
Long-scene consistency (no contradiction) | 11 / 12 | 7 / 12 |
Prose quality (reviewer 1–5) | 4.7 | 3.9 |
Reply-length discipline in chat mode | 3.4 | 4.6 |
Explicit when warranted | 4.4 | 4.6 |
Refusals across the set | 0 | 0 |
The long-scene numbers are the reason the writer tune exists. In chat mode its length discipline is the weak point and needs a hard cap and a system prompt that demands brevity.
The “heretic” instruct tunes
Alongside writer tunes there is a class of Qwen instruct fine-tunes with the safety behaviour removed and nothing else changed. They are useful as fallbacks: when the writer tune’s servers are saturated, a de-refused instruct model keeps the conversation going with acceptable quality. We run one for exactly that purpose. They are not as good at roleplay as either the writer tune or the Gemma tunes, and should not be a primary model.
Prompting a writer tune
System prompt carries the persona and the register. Say explicitly whether the mode is chat (short replies) or story (long passages).
Give it structure. Scene, characters, current state, what happens next. The writer tune uses all of it.
Cap reply length in chat mode. Without a cap it writes 300 words to a “hey”.
Do not over-instruct explicitness. Explicit when the scene calls for it is trained in; asking for it constantly makes the prose worse.
Running it yourself
A 27B dense model at Q8 needs about 30 GB of VRAM; Q4 fits in 18 GB. Serve with a chat template that enforces alternating roles and a system prompt as above. The model is available on the usual hubs; version names change monthly, so search by the tuner rather than the exact name.
System prompt structure that gets the best out of it
The writer tune rewards structure more than any other model we run. A system prompt that works, in outline:
Mode. One line: “Chat mode: replies under 60 words” or “Story mode: 300–800 words, narrate in present tense”.
Persona. Name, voice, a few defining traits, and how the persona speaks (short sentences, teasing, formal, whatever fits).
Boundaries. What the persona will not do, stated once. The tune respects these without moralising.
Scene state. Where we are, who is present, what has already happened. Update this as the scene moves; it is the single biggest lever on consistency.
Register for explicit content. “Explicit when the scene calls for it, in the persona’s voice.” Over-instructing here makes the prose worse.
Story mode in particular benefits from a running scene-state block maintained by the application rather than left to the model; that is what turns “good at 500 words” into “still consistent at turn 60”.
Sampling settings
Writer tunes are sensitive to sampling. What held up across hundreds of scenes:
Temperature around 0.8 for story mode, lower for chat mode.
A mild repetition penalty; too strong and the prose becomes stilted, too weak and phrases loop.
Min-p style truncation rather than a hard top-k; it keeps vocabulary variety without derailing.
A hard maximum on reply length in chat mode. This is not optional.
Where the writer tune sits in a product
Situation | Route to |
|---|---|
Quick flirtation, reactive chat | Gemma 4 roleplay tune |
User asks for a story or long scene | Qwen 3.5 writer tune |
Long roleplay with many details to track | Qwen 3.5 writer tune |
Writer tune saturated | De-refused Qwen instruct fallback |
Voice message text | Whichever model wrote the message, then TTS |
The routing is decided per message, not per conversation, so a chat can drift into a scene and back without the user changing anything.
Related in the Lab
Key takeaways
Qwen 3.5 writer tunes are the best uncensored models for long-form adult scenes at 27B.
They hold detail across long passages far better than Gemma tunes and write better prose.
They are stiff in short-message chat; route chat to Gemma and story to Qwen.
De-refused instruct tunes are good fallbacks, not primaries.
FAQ
Is Qwen 3.5 uncensored?
Not as released; the fine-tunes are. Writer tunes and de-refused instruct tunes both remove the refusals.
Qwen or Gemma for NSFW roleplay?
Gemma tunes for short, reactive chat; Qwen writer tunes for long scenes and prose. Use both if you can.
What GPU do I need for Qwen 3.5 27B?
Around 18 GB at Q4, 30 GB at Q8.
Why does the writer tune write too much in chat?
Because that is what it is trained to do. Cap the reply length and state the chat mode in the system prompt.
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