Himeros V2 27B · Comforting companionship
Warm conversation and comforting companionship
Himeros V2 is a 27B LoRA fine-tune of Qwen3.8-27B-Uncensored, trained with sequences up to 49,152 tokens. Its purpose is to let her talk with me, offer comforting companionship, and help me feel good.
This repository contains the full merged model, not just a LoRA adapter, exported as GGUF precision variants. Each GGUF is standalone: download one file, with no separate adapter or base-model download required.
Model overview
| Detail | Configuration |
|---|---|
| Base model | orcarouter/Qwen3.8-27B-Uncensored |
| Base revision | 404ea47aaa5d8a8b00049c9e9750089aca011ab2 |
| Model size | 27B parameter class |
| Primary language | English |
| Fine-tuning method | LoRA, merged into the base before export |
| LoRA rank / alpha | 16 / 32 |
| Learning rate / schedule | 1e-5 / cosine |
| Training sequence limit | 49,152 tokens (48 × 1,024) |
| Distribution | Single-file GGUF, K and IQ quantizations plus BF16 |
The training mixture combines roleplay conversations, synthetic dialogue and long-form prose. Source training texts are not included in this repository. The training context limit describes the fine-tuning configuration; it does not guarantee uniform quality across an entire 48K conversation.
GGUF collection
The export set comprises the following 23 formats. Consult Files and versions for files currently available; UPLOAD_COMPLETE.json records completion of the full transfer.
| Family | Formats |
|---|---|
| K quantizations | Q2_K, Q2_K_S, Q3_K_S, Q3_K_M, Q3_K_L, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q6_K |
| IQ quantizations | IQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, IQ3_XXS, IQ3_XS, IQ3_S, IQ3_M, IQ4_NL, IQ4_XS |
| Unquantized export | BF16 |
Files follow this naming pattern:
Himeros-V2-27B-<FORMAT>.gguf
These are alternatives, not pieces of one model. Choose a precision that fits your available memory, leaving room for the context cache and runtime overhead. Quantization can affect output quality; this release does not provide a controlled comparison of every format.
Getting started
- Download a single
.gguffrom Files and versions. - Import it into a compatible LM Studio or llama.cpp runtime.
- Use the embedded chat template.
- Set a context length and GPU offload configuration that fit your hardware.
- Supply your character description, scene setup and writing preferences in the system prompt.
Thinking mode
Thinking is not permanently disabled in the GGUF. Whether it is enabled depends on the client and inference backend.
A short local LM Studio test confirmed zero reasoning tokens with Off and reasoning output with Low. This verifies the tested integration, not identical behavior in every client. Exact Low/Medium/High budgets remain backend-dependent.
Export integrity
The original quantization shards were joined losslessly and audited for unchanged tensor names, shapes, types and raw tensor bytes. The transfer workflow compares file SHA-256 against the completed Drive export record and Hugging Face file metadata.
SHA256SUMS.txt provides download checksums once the full transfer completes. These checks verify file integrity, not generation quality. JSON completion records are not required to load a GGUF.
Evaluation and limitations
- No controlled benchmark establishes superiority over the base model.
- Grammar, character continuity, repetition and long-context recall can still fail.
- Quantized variants have not all been independently quality-benchmarked.
- Outputs may contain mature themes, bias or inaccurate information; review them for your use case.
- This is a conversational companion model, not a substitute for human relationships or a source of factual or professional advice.
Attribution and licensing
Thanks to the upstream model authors and the developers of the training and GGUF tooling. Base-model use and redistribution remain subject to the applicable upstream terms. This card does not assert an additional license grant or that third-party training-source rights have been cleared.
Questions
If you have any questions, please DM me.