Macaron-V1 Bolts Four Task-Specific LoRAs Onto a 744B GLM-5.2 Base as a Continual-Learning Architecture

arXiv / HuggingFace Daily Papers (81 upvotes)·medium signalXBlueskyLinkedInCopy link

arXiv 2608.09819 (Aug 10), a 49-page report from 75 authors at Mind Lab, proposes Mixture-of-LoRA (MoL) as the mechanism for open continual learning: the flagship Macaron-V1-Venti pairs a 744B GLM-5.2 base with four separate LoRAs for chat, agent, coding, and GenUI, with a 50B Macaron-V1-Tall variant on Qwen3.6. The interesting part for builders is the surrounding stack they name — MinT post-training platform, MindForge RL framework, LongStraw long-context method, and a component-native GenUI harness called UI4A. It drew 81 upvotes as the second-ranked paper of Aug 11, though the abstract itself withholds numeric benchmark results, so treat the capability claims as unverified until the full report is read.