Himeros 27B
Himeros 27B is an English long-form fictional roleplay and creative-writing model. It was LoRA-tuned from orcarouter/Qwen3.8-27B-Uncensored, merged into the base weights, and exported here as a two-shard BF16 GGUF.
The tuning emphasizes coherent character voice, scene continuity, natural dialogue, grammatical prose, sensory detail, user agency, and sustained multi-turn roleplay.
Adult-content notice: This model is intended only for adults.
Files
Both shards are required and must remain in the same folder with their filenames unchanged.
| File | Drive-reported size |
|---|---|
| Himeros_27B_BF16_01 | 46.55 GB |
| Himeros_27B_BF16_02 | 4.35 GB |
Load Himeros_27B_BF16_02. Compatible GGUF software should locate the second shard automatically.
This is a merged BF16 release, not a standalone LoRA adapter. Quantized releases may be added separately.
Intended use
Himeros 27B is intended for:
- long-form fictional roleplay
- character-driven dialogue and relationship scenes;
- collaborative fiction and scene continuation;
- creative-writing experiments where style and continuity matter.
It is not intended as a factual authority, professional adviser, autonomous agent, or safety classifier. Verify factual claims independently.
Prompting and inference
Use the Qwen chat template supplied by your inference frontend. A clear system prompt should define the character, setting, tone, boundaries, and the rule that the model must not write actions or dialogue for the user unless requested.
For roleplay, disable visible reasoning or thinking in the frontend. If the template exposes an enable_thinking option, set it to false.
Recommended starting settings:
| Setting | Starting value |
|---|---|
| Context | 8192 tokens |
| Temperature | 0.85-1.0 |
| Top-p | 0.90-0.95 |
| Min-p | 0.03-0.08 |
| Repetition penalty | 1.03-1.08 |
These are starting points, not benchmark-optimal values. Lower temperature for tighter continuity; raise it slightly for more variety.
In LM Studio, place both shards in the same model directory, import or rescan the directory, and select the first shard.
Training data
The mixture combined independently reviewed synthetic long-form roleplay examples, a small user-provided romance-dialogue corpus, and filtered public creative-writing and roleplay sources. Public sources included:
- Dampfinchen/Creative_Writing_Multiturn
- angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k
- beyoru/Aesir-Character-CoT-roleplay
- Exxe/literary-roleplay
- jondurbin/gutenberg-dpo-v0.1
Processing included English-language filtering, adult-character constraints, quality checks, near-duplicate removal, grouped train/evaluation splitting, chat-template validation, and response-only masking. A small concise-reasoning anchor was retained for general coherence, while the primary objective remained natural roleplay output rather than visible chain-of-thought.
Dataset inclusion does not transfer ownership of source material. Users must follow the terms and licenses of each upstream source.
Evaluation
No standardized quantitative benchmark result is claimed for this release. Evaluation so far is qualitative and roleplay-focused. A proper comparison should use blinded, identical prompts against the base model and score:
- grammar and readability;
- character consistency;
- scene continuity;
- dialogue naturalness;
- creativity without incoherence;
- respect for user agency;
- repetition and degeneration over long contexts.
Until those results are published, treat claims about improvement over the base as unverified.
Limitations and risks
- The model can hallucinate facts and confidently produce incorrect information.
- It can lose continuity, repeat phrases, over-narrate, or adopt unintended stylistic habits.
- It inherits biases, failure modes, and knowledge limitations from its base model and training sources.
- Fine-tuning for fictional adult roleplay may reduce performance on unrelated factual or coding tasks.
- Prompt wording, sampler settings, context length, and quantization can materially change output quality.
- The model may generate explicit, disturbing, or otherwise objectionable fictional content.
- Training filters reduce risk but do not guarantee safe or policy-compliant output.
Deployers are responsible for appropriate access controls, consent and age safeguards, moderation, privacy protection, and compliance with applicable law.
License and attribution
The repository is marked license: other because this derivative release is subject to the base model's terms and the separate licenses or terms of its training sources. This model card does not grant rights beyond those upstream terms. Review the base model repository and every applicable dataset license before redistribution or commercial use.
Acknowledgements
Built on the Qwen-derived base model released by OrcaRouter, the llama.cpp GGUF ecosystem, Unsloth tooling, Hugging Face infrastructure, and the authors and curators of the listed datasets.