Llama-3.1-8B-Stheno-v3.4

Llama-3.1-8B-Stheno-v3.4

Writes stories and plays characters in multi-turn chat. An 8B Llama 3.1 fine-tune for creative writing and roleplay, aimed at hobbyists who run models on their own machine.

Made by Sao10K on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Llama-3.1-8B-Stheno-v3.4 is an 8 billion parameter model by Sao10K, tuned for creative writing and roleplay. It went through two rounds of fine-tuning, first on multi-turn conversation and instruction data, then on creative writing and roleplay data. The author says it has a different feel from earlier Stheno versions.

Who it suits: It suits writers and hobbyists who run language models locally and want a model built for stories and character chat. The record lists English only.

Industry | Who uses it | How they use it | Example
Creative writing | Fiction writers | Draft scenes and dialogue with a model tuned on creative text | Continue a short story from an opening paragraph
Games and hobby | Roleplay fans | Hold long in-character chats across many turns | Play a ship's navigator in an ongoing story
AI hobby | Local model tinkerers | Try an 8B model with their own sampler settings | Compare the suggested temperature with a lower one

Anyone who needs commercial use or output in another language can skip it, because the licence is non-commercial and the language listed is English.

Where to use it:

• Windows, macOS or Linux with Python, to download the files
• A chat or roleplay front end that can load Safetensors models and apply the Llama 3 Instruct format
• Your own hardware, since the weights are about 16.1 GB in BF16

Use case | Who | What they get back
Story writing | Authors | Narration and scenes in a creative style
Character roleplay | Roleplay fans | In-character replies that hold up over several turns
General assistant questions | Hobbyists | Instruction-style answers, which the card says were part of the tuning

How to set it up: You need a few things first.

• Python and pip
• Disk space for about 16.1 GB, or an estimated 4 GB at 4-bit
• A front end or script that can run an 8B Safetensors model

1. Install the Hugging Face hub tool with pip.
2. Download the model files to your computer.
3. Load them in the front end you use and choose the Llama 3 Instruct prompt format.

pip install -U huggingface_hub
hf download Sao10K/Llama-3.1-8B-Stheno-v3.4

How to use it: The card gives prompting advice rather than commands. Use the Llama 3 Instruct format, and the author says the Euryale 2.1 preset works well. Set temperature and min_p yourself.

• Temperature: the author recommends 1.4
• min_p: the author recommends 0.2
• Experiment, since the author says it behaves differently from earlier versions

Example: You ask it to continue a scene as a ship's navigator and stay in character, using the Llama 3 Instruct format with temperature 1.4 and min_p 0.2. You can expect an in-character reply that carries the exchange forward, the kind of output the roleplay tuning aims at.

For newcomers and experts: If you are new, think of this as a chatty model trained to tell stories and play parts. Download it, load it in a front end that supports the Llama 3 Instruct format, and start with the suggested temperature and min_p.

Under the surface, the tuning data blends human writing with Claude output, and the single-turn examples were redone with new prompts and replies from Opus 3 and Sonnet 3.5. Compared with earlier Stheno sets it holds 55 percent more roleplay and 40 percent more creative writing. The base is Llama 3.1, which the author found harder to wrangle than a Nemo base.

Llama 3.1 8B base model
Stage 1: multi-turn conversation tuning
Stage 2: creative writing and roleplay tuning
Prompt in Llama 3 Instruct format
Model writes the reply

Good to know:

• License: cc-by-nc-4.0, a Creative Commons licence for non-commercial use.
• Cost to run: no API key or paid service named. You supply the hardware, and the compute for training came courtesy of Backyard.ai.
• Data it reaches: only the prompts you give it. The card names no file, web or account access.
• Not for: commercial use under this licence, or languages other than English. The author also calls the Llama 3.1 results less happy than with a Nemo base.

On Hugging Face, 13 November 2026

• Parameters: 8B
• Weights: about 16.1 GB in BF16; about 4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,143
• Likes: 105
• License: cc-by-nc-4.0
• Languages: en
• Quantized versions: 18
• Fine-tunes: 2
• Tags: safetensors, llama

Links

• Model: https://huggingface.co/Sao10K/Llama-3.1-8B-Stheno-v3.4
• Files: https://huggingface.co/Sao10K/Llama-3.1-8B-Stheno-v3.4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Sao10K/Llama-3.1-8B-Stheno-v3.4

Then follow the usage in its model card. Nothing in this kit runs by itself.