Euryale L3.3 70B v2.3 — Model Documentation
Published by Nextbit256 S.L. | Last updated: May 2026 Compliance reference: AI Act Regulation (EU) 2024/1689, Art. 53.1.c and 53.1.d
Model Identification
| Field | Detail |
|---|---|
| Full model name | L3.3-70B-Euryale-v2.3 |
| Developed by | Sao10K (fine-tune author); base model: Meta AI (Meta-Llama-3.3-70B-Instruct) |
| Release date | 2024–2025 |
| Model card | huggingface.co/Sao10K/L3.3-70B-Euryale-v2.3 |
| Technical report | Not published by Sao10K |
| Developer blog | Not published by Sao10K |
Note on catalog identifier: The Nextbit catalog identifier euryale:33-70b refers specifically to version v2.3 of the Euryale series, which uses a Llama 3.3 70B base (not Llama 3.1 or 3.2). Earlier Euryale versions were based on different Llama 3.x variants.
Architecture and Parameters
| Field | Detail |
|---|---|
| Architecture | Dense (decoder-only transformer) — Llama 3.3 architecture |
| Total parameters | 71B |
| Active parameters per token | N/A — dense model (all parameters active) |
| Non-embedding parameters | Not published by Sao10K |
| Layers | Not published by Sao10K (standard Llama 3.3 70B configuration) |
| Attention | Not published by Sao10K (Llama 3.3 70B uses GQA) |
| Native context length | 16,384 tokens (as configured for this fine-tune) |
| Extended context | Not published by Sao10K |
| Precision | BF16 |
License and Commercial Use
License: Meta Llama 3.3 Community License
Euryale L3.3 70B v2.3 is built on Meta-Llama-3.3-70B-Instruct. The Llama 3.3 Community License is a proprietary Meta license — it is not an open-source license such as Apache 2.0 or MIT, and it contains specific conditions.
The Llama 3.3 Community License is distinct from and generally more permissive than the Llama 2 Community License. Key differences include:
- The 700 million MAU threshold applies equally (commercial use without Meta approval requires being below 700M MAU)
- Llama 3.3 derivatives must include "Llama" in the name and display "Built with Llama"
Key terms of the Meta Llama 3.3 Community License:
- Commercial use: Permitted for organizations whose products or services have fewer than 700 million monthly active users (MAU) in the preceding calendar month. Organizations at or above this threshold must request an additional license grant from Meta.
- Attribution requirements: Derivative AI model names must include "Llama" at the start. Providers must display "Built with Llama" prominently.
- Documentation and notice: The license agreement must be distributed with any copy or substantial portion of the Llama materials.
- Acceptable Use Policy: All use must comply with Meta's Acceptable Use Policy.
- Indemnification: Licensees must indemnify Meta against third-party claims arising from their use or distribution.
- Litigation clause: If a licensee sues Meta for IP infringement related to Llama, their license terminates immediately.
Nextbit's verification: Nextbit has reviewed the Llama 3.3 Community License and confirmed that serving Euryale L3.3 70B v2.3 via API for commercial use is permitted, provided Nextbit operates below the 700 million MAU threshold. Nextbit currently operates well below this threshold. The "euryale:33-70b" API identifier satisfies the Llama naming requirement.
Restrictions relevant to users: Users of Nextbit's API who access this model are themselves bound by the Llama 3.3 Community License terms. Users whose own products exceed 700 million MAU must obtain separate Meta approval.
Training Data Summary
This model has two distinct layers: the base model pretraining data and the fine-tune data.
(a) Base model pretraining data — Meta-Llama-3.3-70B-Instruct
Euryale L3.3 70B v2.3 is built on Meta-Llama-3.3-70B-Instruct. The pretraining data for Llama 3.3 is documented by Meta AI. Nextbit does not reproduce that documentation here; for details see the Meta Llama 3.3 model card and Meta AI's published documentation.
Key known facts about Llama 3.3 70B pretraining:
- Approximately 15 trillion tokens (shared across Llama 3.1/3.3 family)
- Web text, code, mathematics, and multilingual content
- Knowledge cutoff: approximately December 2023
(b) Fine-tune data — Sao10K's Euryale v2.3 dataset
Sao10K fine-tuned the base model using LoRA (Low-Rank Adaptation) on three datasets:
- amoral-full-sys-prompt.json — Unalignment / reduced-restriction data (cleaned)
- mimi-superfix-RP-filtered-fixed.json — Roleplay and creative instruction data
- hespera-smartshuffle.json — Hesperus-v2-Instruct data
Fine-tune method: LoRA (lora_r: 128, lora_alpha: 16, lora_dropout: 0.1, RSLoRA enabled); trained using Axolotl v0.5.2 with DeepSpeed ZeRO-3 (BF16), 1 epoch.
Not published by Sao10K:
- Exact size of each dataset
- Sources of the fine-tune datasets (beyond the dataset names listed)
- Whether the fine-tune datasets are publicly available
Languages Supported
Primarily the languages supported by Llama 3.3 70B: primarily English, with multilingual capability. The fine-tune was performed on English-language datasets. Refer to the Meta Llama 3.3 documentation for the full base model language list.
Intended Uses
This model is designed for:
- Roleplaying and character-driven narrative generation
- Creative writing with reduced content restrictions (the fine-tune explicitly includes unalignment data to reduce over-refusals)
- Interactive fiction and storytelling
This is a direct successor to Euryale v2.2, optimized for creative and roleplay use cases. The model is not intended for factual question answering, coding, or scientific reasoning tasks; other models in the Nextbit catalog are better suited for those uses.
Use is subject to Meta's Acceptable Use Policy and the Llama 3.3 Community License.
Systemic Risk Assessment
| Field | Detail |
|---|---|
| Training FLOPs | Not published by Meta AI for Llama 3.3 70B; not applicable for the fine-tune step |
| Relevant base for threshold assessment | Llama 3.3 70B (the pretrained base model); the LoRA fine-tune involves negligible additional compute |
| Estimated FLOPs (Llama 3.3 70B pretraining) | ~1.3 × 10²⁴ (estimate based on 70B parameters × ~15T tokens; formula: 6 × N × D) |
| Exceeds 10²⁵ FLOPs threshold? | No — estimated to be approximately 8× below the systemic risk threshold |
| Note on dense architecture | Llama 3.3 70B is a dense model; all 70B parameters are active per token. |
| AI Office designation | Not designated as a systemic risk model as of May 2026 |
| Art. 55 obligations apply? | No |
The FLOPs relevant for systemic risk threshold assessment are those of the base model (Llama 3.3 70B pretraining). The LoRA fine-tune performed by Sao10K involves a small number of additional trainable parameters and does not materially affect the compute estimate. This estimate is based on publicly available information from Meta AI's Llama 3 documentation. If Meta publishes official compute figures, this section will be updated.
This documentation is published by Nextbit256 S.L. in accordance with Article 53(1)(c) and 53(1)(d) of Regulation (EU) 2024/1689 (AI Act). Nextbit256 S.L. serves this model via its inference API but did not develop or train it. The fine-tune was performed by Sao10K; the base model was developed by Meta AI. All information is sourced from public documentation published by Sao10K and Meta AI. This model is served under the Meta Llama 3.3 Community License.
For questions: [email protected]