Macaron-V1 is an open agent-model family designed for experiential intelligence, enabling learning from real environments and continuing to learn after deployment. The system focuses on two goals: adaptation through recursive improvement of model-harness pairs and collaboration via a Mixture-of-LoRA (MoL) architecture that selects specialist LoRA adapters per user turn.

  • Flagship Macaron-V1-Venti combines a 744B GLM-5.2 base with four LoRAs for chat, agent, coding, and GenUI.
  • Macaron-V1-Tall uses a Qwen3.6-based 50B model for local deployment with the same MoL design.
  • The algorithm integrates Model-Harness Co-design, recursive self-improvement, and the MindForge agentic RL framework.
  • Infrastructure includes the MinT post-training platform and LongStraw long-context RL method.

The report validates the system on Personal Intelligence, GenUI, and general capability benchmarks, though compounding gains from continual learning remain open questions.