mindlab-research/Macaron-V1-Tall

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 22, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Macaron-V1-Tall by MindLab Research is a 35.1 billion parameter Mixture of LoRA (MoL) model built on Qwen3.6-35B-A3B, featuring a 262K context length. It incorporates four specialized LoRA adapters for chat, personal-agent tasks, coding workflows, and Generative UI, with an L0 router dynamically selecting the most suitable specialist per request. This architecture optimizes the model for diverse applications including tool use, repository-level coding, and UI4A Generative UI. It is designed for personal intelligence and complex, multi-domain workflows.

Loading preview...

Macaron-V1-Tall: A Specialized MoL Model

Macaron-V1-Tall, developed by MindLab Research, is a 35.1 billion parameter Mixture of LoRA (MoL) model based on the Qwen3.6-35B-A3B architecture. It stands out with its unique specialist system, comprising four distinct LoRA adapters and an L0 router that intelligently directs user requests to the most appropriate specialist.

Key Capabilities & Architecture

  • Mixture of LoRA (MoL) Architecture: Integrates a Qwen3.6 MoE base with four dedicated LoRA specialists.
  • Specialized Adapters: Features L0 (Chat) for conversational tasks, L1 (Agent) for personal-agent tasks and tool use, L2 (Coding) for code understanding and repository workflows, and L3 (GenUI) for UI4A rendering and UI-driven actions.
  • Dynamic Routing: An L0 router selects the optimal specialist for each new user request, allowing for focused processing within the selected LoRA.
  • Extended Context Length: Supports a substantial 262K context length, enabling handling of extensive inputs.
  • BF16 Checkpoint: Provided as a BF16 checkpoint with four LoRA adapters.

Ideal Use Cases

  • Personal-Agent Workflows: Excels in tasks requiring long-horizon planning and dynamic workflows.
  • Tool Use: Designed for heavy integration and utilization of external tools.
  • Repository-Level Coding: Optimized for complex coding tasks, code understanding, and terminal interactions.
  • Generative UI (GenUI): Capable of UI4A rendering and UI-driven actions, facilitating the creation of user interfaces.

This model is part of the Macaron-V1 family, offering a robust solution for diverse and complex AI applications, particularly those requiring specialized handling across different domains.