KaraKaraWitch/Llama-3.3-MagicalGirl-2.5

KaraKaraWitch/Llama-3.3-MagicalGirl-2.5 Overview

KaraKaraWitch/Llama-3.3-MagicalGirl-2.5 is a 70 billion parameter language model with a 32768 token context length, developed by KaraKaraWitch. It represents an iteration on the MagicalGirl series, specifically modified from MagicalGirl-2 to integrate R1 models, aiming to improve perceived intelligence and reduce 'dumbness'.

Merge Details

This model was created using the mergekit tool, employing the SCE merge method. The base model for this merge was KaraKaraWitch/Llama-3.X-Workout-70B. The merge incorporated several Llama-3.3 based models, including:

• LatitudeGames/Wayfarer-Large-70B-Llama-3.3
• KaraKaraWitch/Llama-MiraiFanfare-3.3-70B
• Black-Ink-Guild/Pernicious_Prophecy_70B
• TheDrummer/Fallen-Llama-3.3-R1-70B-v1
• huihui-ai/DeepSeek-R1-Distill-Llama-70B-abliterated
• SicariusSicariiStuff/Negative_LLAMA_70B

Performance & Characteristics

As of July 3, 2025, the model's UGI-Score is 38.83/100, with specific sub-scores including:

• Unruly: 4.6/10
• Internet: 4.4/10
• Society: 4.6/10
• Willing: 3/10
• NatInt: 31.85/100
• Coding: 24
• Political Lean: -8.7% (Liberalism)

Use Cases

This model is suitable for general language generation and understanding tasks where a 70B parameter model with a large context window is beneficial. Its merged nature, incorporating R1 models, suggests an emphasis on improved reasoning capabilities, making it potentially useful for applications requiring more nuanced responses despite its current benchmark scores.