MindLab Unveils 749B Agent-Optimized Macaron‑V1‑Preview Model

Over the past month, post‑training has emerged as a pivotal engine for boosting large‑model capabilities. In this context, the previously low‑profile Mind Lab open‑sourced Macaron‑V1‑Preview, a 749B‑parameter model (744 + 5 B) built on GLM 5.1 with a 40 B activation size and specifically tuned for deep Agent‑Harness scenarios.

The model was developed using fewer than 300 GPU cards, most of which were not the latest Nvidia chips, resulting in a training compute cost that is under 1% of that required by comparable models of similar scale.

Benchmark evaluations show Macaron‑V1‑Preview achieving state‑of‑the‑art performance across a range of Agent‑focused tasks: it tops LivingBench and VitaBench, matches or exceeds open‑source SOTA on the A2UI protocol and PinchBench, and scores 92.5 on the OpenClaw‑oriented PinchBench, making it the best‑performing open‑source model on that metric.

Technically, the model integrates recent advances such as dense LoRA updates, Agent‑Harness orchestration, and the generative UI (A2UI) protocol. It introduces a native Mixture‑of‑LoRA architecture that allows multiple LoRA adapters to coexist on a single base model, enabling multi‑serve capabilities, dynamic routing, and efficient task‑specific adaptation without retraining the full model.

Mind Lab also released the MinT (MindLab Toolkit) infrastructure, which manages millions of LoRA adapters, reduces model loading latency by nearly tenfold, and supports high‑throughput reinforcement‑learning‑driven post‑training. The system leverages DSA sparse attention and MTP multi‑token prediction to keep training and inference costs low even at the 750B scale.

In a companion 43‑page paper titled “On the Scaling of PEFT,” the team presents a scaling law showing that the accuracy of collaborative model decisions grows logarithmically with the number of participating models, highlighting the benefits of a large population of diverse LoRA adapters for collective intelligence.

Overall, Macaron‑V1‑Preview demonstrates how efficient infrastructure, Mixture‑of‑LoRA, and continuous learning can together enable large‑scale Agent models to evolve rapidly in real‑world settings, marking a practical validation of a new post‑training paradigm.