Mind Lab's Macaron-V1 stacks four LoRA experts on 744B GLM-5.2

Mind Lab's new Macaron-V1 paper is less about topping a leaderboard and more about a compositional bet: freeze a 744B GLM-5.2 base, then swap in one of four specialist LoRA adapters on every user turn — one each for chat, agent, coding, and a GenUI role built on a component-native harness the authors call UI4A.

The flagship is called Macaron-V1-Venti. Alongside it, Mind Lab describes a smaller Macaron-V1-Tall built on a 50B Qwen3.6 base and pitched at local deployment. The wider stack around the models is unusually opinionated for an open release: a post-training platform named MinT, a long-context RL method called LongStraw, and an agentic RL framework called MindForge. Testing spans what the paper labels Personal Intelligence, GenUI, and general capability benchmarks against frontier baselines. It slots into a run of similar architecture posts we've been tracking under agents coverage and open source, where per-turn routing across specialist experts keeps recurring. Motif's 314B MoE with 13.2B active per token landed the same day with a related sparse-activation pitch.

The design choice is worth taking seriously. If per-turn LoRA selection actually works across four quite different task shapes, teams get multi-role behaviour from a single frozen frontier base rather than maintaining and serving four fully fine-tuned copies. That gives Mind Lab a much cleaner update path too — retrain a specialist adapter, leave the 744B base alone.

The caveats sit inside the paper itself: the authors write that compounding gains from continual learning, and the collective-intelligence benefits their pipeline is meant to unlock, remain open questions. What we could pull from the page and abstract does not include concrete benchmark numbers against named frontier models, does not describe weight availability or licensing, and does not break out how much each of the four specialists actually contributes.

The more interesting outcome is not the 744B flagship, which most readers will never serve, but the 50B Tall variant. If it holds up on ordinary local hardware, on-device continual learning gets a lot less academic; if it doesn't, this ends up as an architecturally elegant frozen-base experiment waiting for numbers to catch up.