Macaron-V1 (macaron.im)

Today marks the official release of Macaron-V1, an advanced agent model available in two variants: the 748-billion-parameter Macaron-V1-Venti, which features a 744B base model complemented by four 1B LoRA specialists, and the smaller 35B Macaron-V1-Tall designed for local deployment. This release builds on the Macaron-V1-Preview by introducing significant improvements, such as the GLM-5.2 base model, a smarter agent harness, enhanced long-context capabilities with the LongStraw infrastructure, and expanded evaluation metrics for personal intelligence. These advancements underscore a shift in AI development towards collaborative, specialized models rather than large monolithic structures. The Macaron-V1 innovations illustrate two leading concepts: recursive self-improvement and multi-agent collaboration. Models leverage cumulative experiences to boost their performance while working synergistically with other specialized agents. Key technical details include the MoL architecture that allows for dynamic task handling by routing requests to the most suitable LoRA specialist, and the Harness Context Protocol, which streamlines configurations across training and production. Additionally, the LongStraw framework enables multi-million-token reinforcement learning training, enhancing the model's utility and efficiency. Overall, Macaron-V1 signifies a substantial leap forward for the AI/ML community, offering robust tools for both generative and coding workflows.