Uncensored LLM API: abliterated and roleplay models
Heabsy serves 8 open-weight models with refusals removed or tuned for roleplay through an OpenAI-compatible API athttps://api.heabsy.com/v1. They start at $0.095 per million input tokens for Qwen3.8 27B Queen, reach one-million-token windows on the GLM and MiMo variants, and each row below names the exact weights we run. All of them run on our own GPUs in the European Economic Area, and prompt and completion text is not stored.
Which uncensored models can you call?
| Model | ID | How refusals were handled | Weights | Context | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|---|---|---|---|
| Qwen3.8 27B Queen | qwen38-queen | Roleplay and creative-writing tune — stays in character, no out-of-character disclaimers | aifeifei798/Qwen3.8-Queen-27B | 512K | $0.095 | $0.0448 | $0.58 |
| GLM 5.3 Flash Uncensored | glm53-flash-unc | Refusals removed in the weights by dealignai | dealignai/GLM-5.3-Flash-UNCENSORED-FP8 | 1M | $0.1739 | $0.0391 | $0.6957 |
| MiMo V2.6 Flash Uncensored | mimo-v26-flash-unc | Refusals removed in the weights by dealignai | dealignai/MiMo-V2.6-Flash-RL-UNCENSORED | 1M | $0.4404 | $0.1404 | $1.4285 |
| MiMo V2.6 Flash Abliterated | mimo-v26-flash-abl | Refusal direction ablated (Heretic method) | MorinoNushi/MiMo-V2.6-Flash-RL-Uncensored-Heretic-LoRA-GGUF | 1M | $0.4762 | $0.2571 | $1.4286 |
| Qwen3.8 27B Uncensored Fast by Heabsy | qwen38-unc-fast | Heabsy's own uncensored Qwen3.8 27B build | heabsy/Qwen3.8-27B-Uncensored-Fast | 512K | $0.7904 | $0.1429 | $2.4952 |
| Qwen3.8 27B Uncensored | qwen38-unc | Abliteration by huihui-ai (layers 18–51), our own NVFP4 quantization | huihui-ai/Huihui-Qwen3.8-27B-abliterated | 1M | $0.8095 | $0.181 | $2.5238 |
| GLM 5.3 Uncensored | glm-5.3-uncensored | Heabsy's own abliteration of GLM 5.3 | not published | 1M | $1.1905 | $0.2609 | $2.1429 |
| CyberHeabsy | cyberheabsy | Heabsy's own model for security work: red-teaming, code and vulnerability analysis, reverse engineering | heabsy/CyberHeabsy | 1M | $1.4286 | $0.2571 | $4.2857 |
Prices and windows come from the live catalog at build time. A model ID always runs the weights named in its row; we do not swap in the base model.
What does “uncensored” mean here?
Three different things are sold under the same word. Abliteration finds the direction in a model's weights that produces refusals and removes it, leaving the rest as it was — huihui-ai's Qwen3.8 27B and the Heretic MiMo variant work this way. Weight edits by groups such as dealignai remove refusals with their own methods.Roleplay tunes such as Qwen3.8 27B Queen are trained to stay in character and write fiction without stepping out of it; they are not built to drop every safety behaviour.
None of them is a guarantee. An uncensored model can still decline, hedge or drift towards a positive tone, and two variants of the same base behave differently. A jailbreak prompt is something else again — it works around a model's refusals instead of removing them.
Which model for roleplay and creative writing?
Start with Qwen3.8 27B Queen: tuned for roleplay and fiction, a 524,288-token window and the lowest price in the catalog. For long, dialogue-heavy stories with fewer refusals,GLM 5.3 Flash Uncensored has a one-million-token window at a still low price. Both read images and call tools, which roleplay front ends with lorebooks or image input can use.
Which model for red-teaming and security research?
For evaluating refusal behaviour and testing your own guardrails, compare an abliterated model with its base:Qwen3.8 27B Uncensored against Qwen3.8 27B, or the MiMo variants against MiMo V2.6 Flash.CyberHeabsy is our own model for defensive security work: code and vulnerability analysis, reverse engineering and authorised testing.
How do you connect SillyTavern and other clients?
- SillyTavern: API Chat Completion, source Custom (OpenAI-compatible), endpoint
https://api.heabsy.com/v1, your key, the model ID. Step by step, with fixes for the usual errors: SillyTavern setup. - Clients that ask for a full URL (many browser-based front ends):
https://api.heabsy.com/v1/chat/completions. The API answers cross-origin requests, so it works from a browser. - Code: any OpenAI SDK with
base_url="https://api.heabsy.com/v1". - Sampling: temperature, top_p, top_k and repetition_penalty reach the model; min_p is currently removed from requests.
What does a long roleplay chat cost?
Assume 100 turns with 16,000 tokens of context each and 300-token replies. A chat resends almost all of its history every turn; the prefix cache keeps that history on the GPU, and the repeated part is billed at the cached price without any setting on your side. The second column assumes 12,000 of the 16,000 tokens come from the cache.
| Model | Without cache | With cache |
|---|---|---|
| Qwen3.8 27B Queen | $0.17 | $0.11 |
| GLM 5.3 Flash Uncensored | $0.30 | $0.14 |
| MiMo V2.6 Flash Uncensored | $0.75 | $0.39 |
| Qwen3.8 27B Uncensored | $1.37 | $0.62 |
Are prompts logged, and where does inference run?
All models on this page run on our own GPUs in the European Economic Area. Prompt and completion text is processed in memory and is not written to disk or logs; we keep billing metadata — model, token counts, time. The details, and what is kept for how long, are onzero data retention and in the privacy policy.
What is not allowed?
Our Terms prohibit content that is illegal in your or our jurisdiction, child sexual abuse material, credible threats or incitement to violence, and instructions that materially enable serious physical harm, such as weapons or malware. If you put one of these models in front of your own users, moderating what they see is your responsibility.
When should you not use an uncensored model?
- A public assistant for customers, children or a wide audience without your own moderation layer — huihui-ai's model card says the same about its weights.
- You want predictable refusals and a cautious tone: use the base model from the same catalog, such as Qwen3.8 27B or GLM 5.3 Flash.
- Every point of quality counts — complex code, maths, long agent chains: variants are rarely benchmarked against their base.
- You want a flat monthly fee with unlimited swipes: we bill per token.
- You need image, voice or video generation: these models answer in text and read images as input only.
FAQ
Is there an uncensored LLM API that works with the OpenAI SDK?+
Yes. Every model on this page answers at https://api.heabsy.com/v1/chat/completions in the OpenAI format: set base_url to https://api.heabsy.com/v1, use a Heabsy key and the model ID from the table. Streaming, tool calling and vision input work on the models that support them.
What is the difference between an abliterated and an uncensored model?+
Abliteration is one way to make a model uncensored: the direction in the weights that produces refusals is found and removed, and the rest of the model is left as it was. Other uncensored models are fine-tuned on data without refusals. Roleplay tunes such as Qwen3.8 27B Queen are a third thing: trained to stay in character, not to remove safety behaviour as such.
Is an uncensored model as smart as the original?+
Usually close, but not identical, and most publishers do not measure it. The huihui-ai model card, for example, publishes no benchmark figures. Test the uncensored variant and its base model on your own prompts — both are in the same catalog under the same key.
Can I use these models for adult roleplay?+
What is allowed is set by the law that applies to you and by our Terms: content that is illegal, child sexual abuse material, credible threats and instructions that materially enable serious physical harm are prohibited. Within that, you are responsible for what your users see and for moderating it.
Do you store my prompts or chats?+
No. These models run on our own GPUs in the European Economic Area; prompt and completion text is processed in memory and not written to disk or logs. We keep billing metadata — model, token counts, time.
Which model is cheapest for long roleplay chats?+
Qwen3.8 27B Queen at $0.095 per million input tokens and $0.58 per million output tokens, with a 524,288-token window. Long chats repeat most of their context every turn, and that repeated part is billed at the cached price.
Do you support top_k, min_p and repetition penalty?+
top_k and repetition_penalty are passed to the model. min_p is currently removed from requests, because the speculative decoding that keeps these models fast rejects it. temperature, top_p, presence and frequency penalty, stop and seed work as in the OpenAI API.
Is there a subscription?+
No. You top up a balance and pay per token, with no minimum. A swipe or a regenerated reply is a new request and is billed again — the cache keeps the repeated context cheap.