Llama 3.1 Euryale 70B v2.2: A Roleplay Fine-Tune, Explained

Euryale is not a lab model. It is a community fine-tune by a single author, Sao10K, built on top of Meta's Llama 3.1 70B and aimed squarely at roleplay and creative writing. That framing matters, because the questions you should ask about it are different from the ones you would ask about a general-purpose assistant.

Nobody runs Euryale v2.2 to fix a failing test. People run it because base instruct models write fiction in a flat, cautious register, and a model trained on character chats and prose sounds different. Whether that difference is worth the trade-offs depends on what you are building and, more than most models, on whether you are allowed to build it at all.

What Euryale v2.2 is

The Hugging Face model card describes a single fine-tuning run on Llama 3.1 70B over two epochs, with the datasets kept separate and applied in order rather than merged. First comes multi-turn conversational instruction data, then creative writing and roleplay data alongside some creative instruction sets. The author reports training took about four days on eight A100s.

The data is a mix of human-written and Claude-generated examples. That is common in community fine-tunes and worth knowing: a model partly trained on another model's output tends to inherit some of that model's phrasing habits along with its strengths.

On OpenRouter it is listed as sao10k/l3.1-euryale-70b, text in and text out, with a 131,072-token context window and structured outputs listed as a capability. That figure matches the Llama 3.1 base window; the model card itself does not state a tested length, so treat long-context behaviour as something to verify rather than assume.

What changed from v2.1

v2.2 replaced the Llama 3 base of v2.1 with Llama 3.1, and the author's changelog is unusually candid about the data:

• Multi-turn instruction sets were added to improve coherence across multi-turn conversations.
• Single-turn instruction answers were regenerated with Claude 3.5 Sonnet and Claude 3 Opus.
• Roughly 55% more roleplay examples and roughly 40% more creative writing examples went in.
• New sets targeted system prompt adherence and reasoning about spatial awareness.
• The data was filtered for common errors and stock phrasing, with the author noting that some may have slipped through.

The same card is frank about the base model, describing the Llama 3.1 base as weird compared with Llama 3 and saying the 70B sometimes bugs out, which a regeneration fixes. That honesty is useful. It tells you to budget for retries in any automated pipeline built on it.

Check the licence before anything else

The repository lists the licence as CC-BY-NC-4.0, a non-commercial Creative Commons licence, sitting on top of a base model that carries Meta's own community licence. For a hobby project or private use that is not a problem. For a paid product it is the end of the conversation unless you have a separate arrangement.

Hosted access through an API provider does not change what the weights' licence says about your use; read the provider's terms and the model's licence together. Open weights vs open source covers why downloadable weights and permission to commercialise are two different things, and this model is a clean example of the gap.

Sampler settings matter more than usual

The card recommends temperature 1.2 with min_p 0.2, and the Llama 3 Instruct prompt format. Both deserve attention.

Temperature 1.2 is high by assistant standards, where settings below 1.0 are common. The min_p setting is what keeps it sane: min_p discards any token whose probability falls below a fraction of the most likely token's probability, so the model gets wide choice when several continuations are plausible and narrow choice when one is clearly right. Without min_p, a low temperature tends toward safe, flat prose, and 1.2 risks drifting into incoherence, so test the pair together rather than either knob alone. Temperature, top-p and sampling explains the mechanics if the knobs are unfamiliar.

Not every API exposes min_p. If yours does not, start lower, around 0.9 to 1.0 with top_p, and adjust by reading output rather than by copying numbers.

Euryale v2.2 or the Llama 3.3 version

The author later released a successor on Llama 3.3 70B Instruct, covered in the Llama 3.3 Euryale 70B guide. The card for that model calls it a direct replacement for v2.2. In practice:

• Pick v2.2 if you have prompts and character cards already tuned against it and they work. Prose style is the product here, and a new version changes the feel even when it is nominally better.
• Pick the Llama 3.3 version for new work. It is the newer release, it was listed cheaper on OpenRouter at the time of writing, and its repository carries a Llama licence tag rather than a non-commercial one.

On cost, v2.2 was listed on OpenRouter at $0.85 per million tokens for both input and output at the time of writing. Flat pricing like that makes long chat histories expensive on the input side, since every turn resends the conversation. If you want plain Llama 3.1 behaviour for comparison, the Llama 3.1 70B Instruct guide covers the base it was tuned from.

How to evaluate a creative model

Benchmarks are close to useless here. There is no published score that predicts whether a character stays in voice over forty turns. Build your own small test instead:

1. Write three scenes that represent your actual use: one dialogue-heavy, one descriptive, one with several characters in a shared space.
2. Run each for twenty turns with the recommended sampler, then with your platform's defaults.
3. Score for consistency of character voice, whether it remembers established facts, and repeated phrases across turns.
4. Count how often you had to regenerate. That number is your real cost multiplier.

If the model holds voice and the licence fits your use, it does its job. If you need commercial rights, stop at the licence section and look at the successor.