Sao10k L3 8B Lunaris is an 8 billion parameter language model built on the Llama 3 architecture through a deliberate merge of several specialized Llama 3 variants. Rather than training from scratch, the creator selected and combined models with complementary strengths—including variants trained for roleplaying and storytelling alongside those contributing general knowledge and reasoning capabilities. This merger strategy aims to overcome the typical tradeoff in merged models, where creative generation often comes at the cost of logical consistency. The result is designed to serve both as a general-purpose assistant and as a capable tool for narrative and character-driven interactions, offering the versatility of a merged model without sacrificing coherence in either direction.

The model was constructed using the ties merge method with Meta-Llama-3-8B-Instruct as the foundation, applying careful density and weight parameters to each contributing model based on extensive experimentation. Contributing models include maldv/badger-iota-llama-3-8b for general reasoning, and roleplaying-specialized variants like crestf411/L3-8B-sunfall-v0.1 and Hastagaras/Jamet-8B-L3-MK1, each bringing distinct strengths to the final blend. The creator describes values chosen from long-running personal experimentation since the Llama 2 merging era, tweaked to fit this specific recipe. With open weights available and a chat template using Llama-3-Instruct formatting, the model supports structured outputs and offers users a balance between creative generation and logical coherence that its creator considers an improvement over earlier merged models like Stheno v3.2.