Mind Lab Releases Macaron-V1 Model Family Built on Qwen 3.6
Mind Lab has introduced the Macaron-V1 agent model family, which includes variants built on the newly released Qwen3.6-35B-A3B base model. The family features a "Tall" 35B variant and a larger "Venti" 748B variant utilizing a Mixture-of-LoRA (MoL) architecture. This release showcases the rapid adoption of the new Qwen 3.6 open-weight models for downstream agentic applications. By leveraging Mixture-of-LoRA, it provides a resource-efficient way to scale large agent models for general life scenarios. The Macaron-V1 family utilizes a model-harness co-design, with the preview version (Macaron-V1-Preview) being a 749B model post-trained from GLM5.1 using MinT. The architecture combines a base model with multiple LoRA adapters (such as a 744B base with five 1B LoRA adapters) to optimize performance.
## BACKGROUND
Qwen 3.6 is the latest iteration of Alibaba's open-weight large language model series, with Qwen3.6-35B-A3B being one of its initial variants. Mixture-of-LoRA (MoL) is an architectural approach that combines multiple Low-Rank Adaptation (LoRA) modules, allowing models to specialize in different tasks efficiently without retraining the entire massive base model.