A new open model family, Macaron-V1, introduces an architecture designed to learn from experience in real environments and continue learning after deployment — a step away from static models that are frozen after training. The system is built around two goals: adaptation through recursive improvement of versioned model-harness pairs, where experience from one configuration is evaluated under an external contract and used to construct its successor, and collaboration via the Mixture-of-LoRA (MoL) architecture, which freezes a base model, composes specialized LoRA adapters, and selects one LoRA per user turn. The flagship model Macaron-V1-Venti combines a 744B GLM-5.2 base with four LoRA adapters for chat, agent, coding, and GenUI; the Qwen3.6-based Macaron-V1-Tall (50B) uses the same design for local deployment. Read the full report.
For decision-makers, the key point is not the technical details but the implication: Macaron-V1 is designed to support continual learning through extensible LoRA specialists, meaning the system can be updated and improved without retraining the entire model. This makes it potentially attractive for organizations that need models that can adapt to new data or tasks after implementation — without starting from scratch. At the same time, it is worth noting that the authors themselves assess that the compounded gains from continual learning and collective intelligence are still open questions, indicating that the technology is not yet mature for broad enterprise adoption.
Conclusion: Macaron-V1 represents a step toward models that can learn after deployment, but with open questions about the overall gains. For decision-makers, it is worth following the development, but not necessarily investing today.
