Qwen3.8-27B landed with an Apache 2.0 licence, vision input and a 262K context, and the "uncensored" versions were on Hugging Face the next day. Most of them are abliterations, not fine-tunes. Here is what that distinction means for adult roleplay, and when a real successor to the 3.5 writer tunes is likely.
Short answer: Qwen3.8-27B is the best base model at its size and it is fully open. The uncensored builds that appeared within a day of the 14 August release are abliterated: the refusal direction removed from the weights, nothing added. That makes them useful right now as de-refused instruct models, and not yet a replacement for the Qwen 3.5 writer fine-tunes that MyBabes runs for long-form roleplay, which were trained on adult fiction rather than merely un-refused. Expect proper 3.8 roleplay tunes within one to two months; we will switch when they beat 3.5 in our evaluation.
What shipped
Qwen3.8-27B | Qwen3.8-Max | |
|---|---|---|
Released | 14 August 2026 | 12–13 August 2026 |
Parameters | 27.8B dense | 2.4T MoE, 95B active |
Licence | Apache 2.0 | Custom |
Input | Text, images, video | Text (vision and 1M context stripped) |
Context | 262,144 tokens | Large |
Runs locally | Yes, quantised on a high-end workstation | No |
The 27B is the one that matters for anyone running their own models. Alibaba reports large gains over Qwen3.6-27B in agentic coding, computer use and vision-language work at the same size.
Abliteration versus fine-tuning
The distinction decides whether a model is good for roleplay or merely willing.
Abliterated build | Roleplay / writer fine-tune | |
|---|---|---|
What changed | Refusal direction orthogonalised out of the weights | Instruction layer retrained on roleplay and adult fiction |
Refuses adult content | No | No |
Stays in character | As well as the base, which is: not very | Trained for it |
Escalation pacing | None; it answers what it is asked | Trained |
Prose in long scenes | Assistant-like | Trained |
Available for 3.8 | Yes, since 15 August (FP8, GGUF, several variants) | Not yet |
The abliterated 3.8 builds are exactly what we described as “heretic” fallbacks in the Qwen 3.5 analysis: they keep a conversation going without refusals, and they are not persona models. Prompt-level “uncensored packs” that only mask refusals in the system prompt are a third, weaker category, and several of the search results for “Qwen 3.8 uncensored” are those.
What we did with it
We ran the abliterated 27B in FP8 through the chat portion of our roleplay evaluation, against the Qwen 3.5 27B Writer tune and the Gemma 4 gendered tunes.
Refusals: none, as expected.
Character consistency: below both tuned models; it drifts back to an assistant register within a few turns unless the system prompt is heavy.
Knowledge and coherence: noticeably above 3.5 in long contexts, which is the promise for future tunes.
Explicit prose: competent, flat; reads like an assistant describing a scene.
As a base for the next writer tune it looks excellent. As a drop-in replacement today, it is not.
What to expect and when
The 3.5 generation’s roleplay tunes appeared roughly four to eight weeks after the base weights. The same tuners are active, the licence is friendlier than 3.5’s, and the 262K context is attractive for long scenes. A reasonable expectation is a credible 3.8-27B writer tune by October 2026. MyBabes will evaluate candidates against the current 3.5 tune on the same 40-scenario set and switch when one wins; the Gemma 4 tunes stay in place for short-message chat regardless.
Running it yourself
A 27B model at Q4 needs about 18 GB of VRAM, FP8 around 30 GB. The abliterated builds serve on the same vLLM and llama.cpp paths as the official release; the GGUF variants are the easy route on a single consumer card. Use a strong system prompt for persona work; the model needs it more than a tune would.
Related in the Lab
Key takeaways
Qwen3.8-27B is Apache 2.0, multimodal, 262K context, and the strongest open base at its size.
The day-one “uncensored” versions are abliterations: no refusals, no roleplay training.
They are good fallbacks, not replacements for the 3.5 writer tunes.
Expect real 3.8 roleplay tunes within one to two months; we will switch when one wins the evaluation.
FAQ
Is Qwen3.8 uncensored?
The official model is safety-tuned. Abliterated community builds with refusals removed appeared on 15 August 2026.
What is the difference between abliterated and fine-tuned?
Abliteration removes the refusal behaviour from the weights without adding skills. A fine-tune retrains the model on new data, which for roleplay means persona, pacing and prose.
Should I switch from Qwen 3.5 to 3.8 for NSFW roleplay?
Not yet. Wait for a writer tune on 3.8; use the abliterated build only as a fallback.
Published: 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
Editorial team
The MyBabes Lab team tests AI image, video and chat models for adult use cases: what they allow, what they block, and how to get the best output. Every note here is based on hands-on runs, not vendor marketing.
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