Llama 3.1 Euryale 70B v2.2

Llama 3.1 Euryale 70B v2.2 is a community finetune by Sao10K, published on Hugging Face as Sao10K/L3.1-70B-Euryale-v2.2. It starts from Meta's Llama 3.1 70B and is tuned for roleplay and creative writing. According to the model card, it went through a single-stage finetune of 2 epochs, with the data applied in two ordered parts: multi-turn conversational instruct data first, then creative writing, roleplay and creative instruct data. The data mixes human-written and Claude-generated examples.

Key features

• Changes from v2.1: per the card, it adds multi-turn instruct data, replaces single-turn instruct data with Claude 3.5 Sonnet and Claude 3 Opus outputs, removes c2 samples, and adds more roleplay and creative-writing data plus data for system-prompt adherence and reasoning or spatial awareness.
• Training: about 4 days on 8x A100 GPUs, 2 epochs, datasets applied in sequence rather than merged.
• Prompt format: use the Llama 3 Instruct format; the author says the Euryale 2.1 preset works well and recommends temperature 1.2 with min_p 0.2.
• License: the card lists cc-by-nc-4.0, which is non-commercial. Also review the upstream Llama 3.1 license terms.
• Author's caveat: the card says it is fine for roleplay but may be less uncensored zero-shot.

Use cases

• Roleplay and character chat.
• Creative writing and storytelling drafts.
• Experiments comparing creative finetunes of the same Llama 3.1 70B base.

Pricing and availability

The weights are free to download from Hugging Face under cc-by-nc-4.0, so commercial use is not covered by the card's license. At the time we checked on 2026-10-29 the card showed 70 likes and about 775 downloads in the last month, and no inference provider hosting it. A 70B model needs substantial GPU memory or a quantized build. The card states no context length.

Getting started

1. Download the model from the Hugging Face repo, or use a quantized community build.
2. Format prompts with the Llama 3 Instruct template.
3. Start with temperature 1.2 and min_p 0.2, as the author recommends, then tune.
4. Confirm the license fits your use before building on it.

FAQ

What is Euryale v2.2?

A Llama 3.1 70B finetune by Sao10K for roleplay and creative writing.

What prompt format should I use?

The Llama 3 Instruct format, per the card. The Euryale 2.1 preset is said to work well.

Can I use it commercially?

The card lists cc-by-nc-4.0, a non-commercial license, so check carefully before commercial use.

What changed from v2.1?

More multi-turn and roleplay data, Claude 3.5 Sonnet and Claude 3 Opus outputs for instruct data, and data for system-prompt adherence. See the v2.1 entry for the earlier version.

Alternatives

• Llama 3 Euryale 70B v2.1: the earlier Euryale version.
• Llama 3.1 70B Instruct: Meta's own instruct model on the same base.
• MythoMax 13B: a smaller Llama 2 roleplay merge.

Conclusion

Euryale v2.2 is a documented community finetune for roleplay and creative writing, with clear training notes but a non-commercial license. Read the model card before use and see related options in the models category.