Release2026-08-03
Chinese startup Mind Lab released and open-sourced Macaron-V1 on 21 July, a model built by attaching four small trainable add-on modules to a frozen GLM-5.2 base so only four billion of its 748 billion settings are changed. The company reports best-in-class scores on six of twelve benchmark tests, all self-published. A lighter 50-billion-parameter version built on Qwen 3.6 is intended to run locally.
What changed
Adapting a very large model to a use case meant expensive full retraining or complex prompting scaffolds.
What it unlocks
Running an open-weight model whose behaviour is tuned by swapping small add-on modules rather than retraining the whole model.
- 748B params, 4B trained
- SOTA in 6 of 12 tests
- 2M token context window
- USD 10M ARR
What you need to act on it
- self-hosting or the MinT platform
- GPU capacity for LoRA training
Sources