MindLab Research Serves Up Macaron-V1-Preview-749B Personal AI Agent

    
        By vramkickedin    
     | 
    
            June 22, 2026 at 11:14 pm        
    
     | 
    
        2 min read

MindLab Research has released Macaron-V1-Preview-749B, a 749-billion-parameter AI model built to serve as a personal agent that can use tools and generate dynamic user interfaces. The system combines a massive 744-billion-parameter base model with five specialized adapters, each trained for a different type of task like planning, coding, or creating visual cards and forms. A built-in router automatically switches between these adapters depending on what the user needs, all within a single conversation.

The model was post-trained from the GLM-5.1 foundation using the MinT system to handle real-world scenarios where user intent, tools, and environment keep changing. It operates with a purpose-built harness that manages routing, memory, and tool calls, aiming to keep the agent coherent across multi-turn interactions. This preview release arrives under the MIT license and lives in one Hugging Face repository containing all base weights and adapter files.

Mixture-of-LoRA for personal agent tasks

Key Features

Combines 744B base model with five specialist LoRAs.
Explicit router-tool design for task switching.
Handles personal planning, search, and calendars.
Supports coding, terminal, and shell workflows.
Generates interactive UI elements via A2UI.
202,752 token context window.
Released under MIT license.

Developers and researchers building local personal assistants can use this model to power agents that understand daily life tasks, remember user preferences, and display answers as interactive cards or forms. The preview is designed for testing agent behavior, tool grounding, and generative UI in contained environments. Because full performance relies on the specific routing harness, anyone experimenting should follow the official setup to reproduce benchmark results.

Research preview with harness dependency

Macaron-V1-Preview-749B is a model-and-harness release, meaning its intended behavior depends on a companion serving system that orchestrates adapter routing and tool execution. If the harness is swapped out or simplified, the developers warn that benchmark scores and task accuracy will change. Deployments must also guard private user state and require explicit confirmation before any external write action like booking, messaging, or purchases.