EVA-LLaMA-3.33-70B-v0.1 is a 70 billion parameter specialist built upon the Llama-3.3-70B-Instruct foundation, developed by Kearm, Auri, and Cahvay. Rather than training from scratch, the team used a DELLA linear merge approach across previous checkpoints to construct this model with specific behavioral goals: reducing overfitting, improving comprehension and recall across its 32,768 token context window, enhancing output stability, and minimizing repetitive patterns in generated text. The result is a model explicitly tuned for roleplay and creative storywriting, where maintaining character consistency and narrative flow over long exchanges matters most.

The model was fine-tuned on a diverse mixture of both synthetic and natural datasets, including Celeste 70B 0.1, Kalomaze's Opus Instruct 25k, Gryphe's ChatGPT-4o-WritingPrompts, and Epiculous's Synthstruct and SynthRP datasets. This expanded data mixture is designed to give the model a richer creative flavor compared to its base, supporting more varied and engaging output. The model has already proven its value as a storytelling component in subsequent merges, where it was combined with scene-description specialists like EURYALE-v2.3 and prose-enhancement models such as Anubis-v1 to create more sophisticated narrative tools. Developers recommend specific sampler settings—temperature at 1, min-P at 0.05, and repetition penalty at 1.03—for optimal creative generation results.