Macaron-V1 - MindLab's post-training multi-capability model system

What is Macaron-V1

Macaron V1 is a post-training multi-capability model system launched by MindLab, based on GLM 5.2 The base model uses reinforcement learning to perform LoRA fine-tuning on approximately 0.5% of core parameters. The system pioneered the Mixture-of-LoRA (MoL) architecture, mounting four independent experts: Chat, Agent, Coding, and UI, activated on demand by a Router, completely eliminating capability interference. Macaron V1 offers the flagship Venti and lightweight Tall versions, covering dialogue, agents, coding, and interactive UI generation, with a measured generation speed of 320 characters/second.

Macaron-V1's main features

• MoL post-training architecture: The base model is frozen, only 0.5% of core parameters are trained, and four independent LoRA experts are isolated and optimized without interfering with each other. Adding new capabilities only requires mounting another LoRA.
• Dual model variants: Flagship Venti (748B, based on GLM 5.2, native 1M context extended to 2M); lightweight Tall (35B, based on Qwen 3.6, optimized for local deployment).
• Three supporting systems: UI4A universal generative UI architecture, rendering interactive views; REPL Harness persistent Python environment, where the model writes its own tools and promotes them to global tools; MinT self-developed training platform, transmitting only Adapters, supporting million-level directories and trillion-parameter scale.
• Self-built evaluation benchmarks: Macaron ChatBench supports long-context honest alignment, and LivingBench provides a dynamic sandbox environment to verify continuous understanding and task replanning.
• Extreme generation speed: API measured at 320 characters/second, completing a 10,975-character long text in 34.3 seconds.

Macaron-V1's technical principles

• MoL (Mixture-of-LoRA) architecture: Freeze all weights of the GLM 5.2 base model, extract only about 0.5% of the core parameters to build four independent LoRA expert modules, each specializing in one of the four major capability directions: Chat, Agent, Coding, and UI. Each expert is independently optimized in an isolated adaptation space, completely eliminating the common "press down the gourd only to have the ladle float up" interference problem in joint multi-capability training. User input is routed to the corresponding expert via an explicit Router Tool, achieving on-demand activation and efficient resource utilization.
• LongStraw long-context extension: Based on the self-developed LongStraw technology, the native 1M context window of GLM 5.2 is extended to 2M. The core mechanism is to reuse shared prompts as persistent states, avoiding the overhead caused by repeated computation in long sequences, greatly reducing the resource consumption and latency of long-context inference, and enabling the model to remain efficient and stable when processing ultra-long documents and complex multi-turn dialogues.
• MinT distributed training platform: MindLab's self-developed post-training infrastructure. Its core design breaks through the bottleneck in traditional distributed training where Actor and Learner need to transmit complete weights, instead transmitting only lightweight Adapters. This design naturally fits the MoL architecture, supports million-level Adapter directory management, and can cover end-to-end reinforcement learning training for models at trillion-parameter scale, significantly reducing communication bandwidth and storage pressure.
• Real-environment-driven reward functions: Unlike training paradigms that rely on manual annotation or static benchmarks, Macaron emphasizes reward signals from real product environments. The team built two dynamic evaluation benchmarks, Macaron ChatBench and LivingBench: the former examines honest alignment and task completion capabilities in long contexts, while the latter verifies the model's continuous understanding, information verification, and task replanning capabilities in a sandbox containing dynamic noise, dynamic environments, and dynamic users, ensuring the model continuously evolves in real interactions.

How to use Macaron-V1

Macaron V1 is in beta testing and can be accessed via the macaron gateway https://mintcn.macaron.xin to call the API.

Core Advantages of Macaron-V1

• Capability isolation without interference: The MoL architecture allows each expert to be optimized independently, completely avoiding the performance trade-offs that occur when training multiple capabilities together.
• Extremely low-cost expansion: Adding new capabilities only requires mounting a new LoRA, with no need to retrain the entire model, allowing flexible combination.
• Ultra-long context support: LongStraw technology extends the context to 2M and reduces the overhead of long sequences through state reuse.
• Evolution in real environments: With a self-built dynamic evaluation benchmark, the model continuously iterates in an environment that simulates real noise and dynamic users.
• Leading generation speed: Measured at 320 characters/second, far exceeding mainstream models on the market, with continuous and stable output after the first token response.

Comparison of Macaron-V1 with Similar Competitors

Application Scenarios of Macaron-V1