749B total (744B base + 5B LoRA adapters) / 40B active
202,752 tokens
MIT
Official open-weight preview release in Mind Lab's Macaron V1 series, positioned as a personal-agent-focused post-trained model built on top of GLM-5.1 rather than a from-scratch frontier pretraining run. Release materials describe Macaron-V1-Preview-749B as a Mixture-of-LoRA architecture that combines a 744B GLM-5.1 base with five specialist 1B adapters for general chat, personal-agent workflows, coding and shell tasks, generative UI, and computer-agent behavior, with roughly 40B activated parameters at inference. The launch framing emphasizes explicit specialist routing for agent harness scenarios instead of a single opaque post-trained policy, aiming at multi-turn personal assistance where user goals and constraints evolve across the conversation. Mind Lab's related Macaron-A2UI research describes the broader 754B-class Macaron line as trained with parameter-efficient LoRA-based supervised fine-tuning followed by reward-driven reinforcement learning for personal-agent and generative-UI tasks, and public coverage of the Preview release highlights strong results on LivingBench, VitaBench, A2UI evaluation, PinchBench, and general math and coding tasks relative to contemporary open models. The release is also presented as unusually compute-efficient for its scale, with Mind Lab claiming development on fewer than 300 GPUs while keeping the checkpoint open under an MIT license.