Roleshade 1B
Status: 🚧 Weights pending release — expected in 2–3 weeks.
Roleshade 1B is a 1-billion-parameter model built by Lightshade for roleplay and creative writing. It is designed to run well on hardware people actually own — phones, laptops, and CPU-only machines — without giving up long-context conversations.
This repository is a placeholder. Model files, quantizations, and usage examples will be published here on release.
Highlights
- Purpose-built for character work. Trained for persona consistency, scene memory, and prose that resists the repetition collapse common at this parameter count.
- Fast prefill on CPUs and phones. Roleshade 1B uses a custom architecture that addresses the prefill bottleneck that makes long-context local models painful on non-GPU hardware. Loading a long chat history doesn't mean waiting on a spinner.
- Long context, anywhere. Deep, persistent roleplay sessions on a laptop, desktop, or phone.
- Local by default. Runs entirely on your own machine. Nothing leaves the device.
Intended Use
Primary: Interactive roleplay, character simulation, collaborative fiction, and creative writing assistance.
Out of scope: Factual question answering, coding assistance, math, retrieval-grounded work, safety-critical applications, or any use where the model's output is treated as authoritative. Roleshade 1B is a creative model and will confidently produce fiction.
Benchmarks
⚠️ Important: The scores below are from the base instruct model, not from Roleshade 1B itself. Roleshade 1B is heavily fine-tuned for roleplay and creative writing, and will not necessarily reproduce these results. Benchmark numbers for the released roleplay model will be published separately, alongside creative-writing-specific evaluations.
| Benchmark | Score |
|---|---|
| BBH | 78.7 |
| GSM8K | 59.1 |
| GPQA-Diamond | 39.5 |
| MBPP | 51.2 |
| MMLU | 58.5 |
| MMLU-Pro | 52.7 |
| MMLU-Redux | 82.4 |
| IFEval | 88.4 |
Usage
Code samples will be added when weights are published.
# Available on release
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "lightshade/roleshade-1b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
Roleshade 1B will also be available directly on LSAI at launch — one click, no setup required.
Recommended sampler settings and prompt/character-card formatting will be documented here at release.
Limitations
- At 1B parameters, world knowledge is limited. Expect confident hallucination on factual topics.
- Optimized for creative output; instruction-following outside of roleplay contexts may be less reliable than the base model's benchmarks suggest.
- Long-session coherence is strong for the size class but not unlimited.
- English-first. Other languages are untested.
Responsible Use
This model is intended for adult creative fiction and roleplay. Users are responsible for how they deploy it and for complying with applicable law and the terms of any platform they run it on. Outputs are fiction and should not be relied on as advice of any kind.
License
Apache 2.0 (subject to confirmation at release).
Citation
@misc{roleshade1b,
title = {Roleshade 1B},
author = {Lightshade},
year = {2026},
url = {https://huggingface.co/LightshadeML/Roleshade-1B}
}