Skttttt/Himeros_27B_IQ4XS

Himeros 27B IQ4_XS

Himeros 27B IQ4_XS is the compact llama.cpp and LM Studio release of Himeros 27B, 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 in BF16, and then quantized to IQ4_XS.

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. Its training mixture includes explicit fictional content involving adult characters. Do not use it to depict minors or age-ambiguous characters, or for non-consensual real-world exploitation.

File

File	Quantization	Approximate size
Himeros_27B_IQ4XS	IQ4_XS	14.26 GiB / 15.31 GB

This is a single-file, merged GGUF release. It is not a standalone LoRA adapter or training checkpoint.

IQ4_XS is a lossy inference quantization. It requires substantially less storage and memory than BF16, but it can lose some writing subtlety, accuracy, and consistency compared with the merged BF16 model.

Keep the original adapter or BF16 merge if you intend to continue training or create additional quantizations.

The IQ4_XS export was produced through Unsloth and llama.cpp using the Qwen3.8-27B importance matrix published by bartowski/Qwen3.8-27B-GGUF.

Download

Using the Hugging Face CLI:

hf download Skttttt/Himeros_27B_IQ4XS --local-dir .

Use a recent version of LM Studio or llama.cpp with support for the Qwen3.8 architecture.

LM Studio

1. Download the GGUF through LM Studio, or place it in an LM Studio model directory.
2. Rescan the model directory if necessary.
3. Select Himeros_27B_IQ4XS.
4. Begin with an 8192-token context.
5. Offload as many layers to the GPU as your VRAM allows.
6. Disable visible reasoning or thinking for roleplay.

The 14.26 GiB file size is not the complete runtime-memory requirement. Model buffers, context/KV cache, and backend overhead require additional VRAM or system RAM.

Partial CPU/RAM offloading is supported but will reduce generation speed.

llama.cpp example

llama-cli -m Himeros_27B_IQ4XS.gguf -cnv -c 8192 -ngl 999 --temp 0.9 --top-p 0.95 --min-p 0.05 --repeat-penalty 1.05

If full GPU offloading does not fit, reduce -ngl or use your frontend’s automatic GPU-offload setting.

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 embedded in the GGUF or supplied by a compatible frontend. Do not apply an unrelated manual prompt template.

For roleplay, disable visible reasoning or thinking. If the template exposes an enable_thinking option, set it to false.

A clear system prompt should define:

• the character and setting;
• the desired tone and writing style;
• relevant boundaries;
• established scene details;
• the rule that the model must not write for the user.

For English-only roleplay, consider adding:

Write only in natural, grammatical English unless the user explicitly requests another language.

Stay in character and preserve established scene details.

Never write the user's dialogue, thoughts, choices, or actions unless explicitly requested.

Respond directly without analysis or think tags.

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 rather than benchmark-optimal settings.

Lower the temperature for tighter continuity. Raise it slightly for more variation. If the model becomes repetitive, first adjust the prompt or recent context before applying a strong repetition penalty.

Training data

The training 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 training/evaluation splitting;
• chat-template validation;
• response-only masking.

A small concise-reasoning anchor was retained for general coherence, while the main 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 has been qualitative and roleplay-focused.

A proper comparison against the base model should use blinded, identical prompts and evaluate:

• grammar and readability;
• character consistency;
• scene continuity;
• dialogue naturalness;
• creativity without incoherence;
• respect for user agency;
• repetition and degeneration over long contexts.

Until comparative results are published, treat claims of improvement over the base model as unverified.

IQ4_XS is also expected to lose some accuracy and subtlety compared with the merged BF16 release.

Limitations and risks

• The model can hallucinate facts and confidently provide 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.
• Roleplay fine-tuning may reduce performance on unrelated factual, reasoning, or coding tasks.
• IQ4_XS trades some output quality for lower storage and memory requirements.
• Prompt wording, sampler settings, context length, and hardware offloading can materially affect output.
• 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 using:

• the Qwen-derived base model released by OrcaRouter;
• the llama.cpp GGUF ecosystem;
• Unsloth training and export tooling;
• Hugging Face infrastructure;
• Bartowski’s Qwen3.8-27B importance matrix;
• the authors and curators of the listed datasets.