Hugging Face has released Macaron-V1, an open agent-model family aimed at achieving experiential intelligence—the ability to learn from real-world interactions and continue improving after deployment. The system is built around two core goals: adaptation through recursive improvement of versioned model-harness pairs, and collaboration via a Mixture-of-LoRA (MoL) architecture that freezes a base model and dynamically selects a specialist LoRA adapter for each user turn.
The flagship Macaron-V1-Venti combines a 744B-parameter GLM-5.2 base model with four LoRA adapters specialized for chat, agent tasks, coding, and GenUI. For local deployment, Macaron-V1-Tall (50B) uses the same design but is based on Qwen3.6. This co-designed system spans architecture, algorithms, and infrastructure, enabling continual learning through extensible LoRA specialists.
Key algorithmic components include Model-Harness Co-design and a recursive self-improvement loop, featuring the UI4A component-native GenUI harness, a stateful action substrate, a versioned HCP contract, and the MindForge agentic RL framework. Supporting infrastructure comprises the MinT post-training platform, the LongStraw long-context RL method, and stability techniques for sparse MoE and DSA base models.
Macaron-V1 was evaluated on Personal Intelligence, GenUI, and general capability benchmarks against frontier baselines. The results validate the current system, while the authors note that compounding gains from continual learning and collective intelligence remain open questions.