Macaron-V1 - A post-trained multi-capability model system from MindLab

What is Macaron-V1?

Macaron V1 is a post-trained multi-capability model system launched by MindLab. Based on the GLM 5.2 platform, it uses reinforcement learning to fine-tune approximately 0.5% of core parameters using LoRA. The system features a unique Mixture-of-LoRA (MoL) architecture, mounting four independent experts: Chat, Agent, Coding, and UI, which are activated on demand by the Router, completely eliminating capability interference. Macaron V1 offers the flagship Venti version and the lightweight Tall version, covering dialogue, agent, programming, and interactive UI generation, with a tested generation speed of 320 characters per second.

Main features of Macaron-V1

• MoL post-training architectureThe base is frozen, only 0.5% of the core parameters are trained, and four independent LoRA experts are optimized in isolation so that they do not interfere with each other. Adding new capabilities only requires adding another LoRA.
• Dual-model variantFlagship 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 major supporting systemsUI4A general generative UI architecture, rendering interactive views, REPL Harness persistent Python environment, model self-writing tools and upgrading them to global tools, MinT self-developed training platform, only passing Adapter, supporting millions of directories and trillions of parameters.
• Self-built evaluation benchmarksMacaron ChatBench supports long context honest alignment and LivingBench dynamic sandbox environment for verifying continuous understanding and task replanning.
• Extreme generation speedAPI test results show a speed of 320 words per second, completing a 10,975-word article in 34.3 seconds.

Macaron-V1 Technical Principles

• MoL (Mixture-of-LoRA) architectureThe entire weights of the GLM 5.2 base are frozen, and only about 0.5% of the core parameters are extracted to build four independent LoRA expert modules, specializing in Chat, Agent, Coding, and UI capabilities respectively. Each expert is independently optimized in an isolated adaptation space, completely eliminating the "gourd-like" mutual interference problem commonly encountered in multi-capability joint training. User input is explicitly routed to the corresponding expert via the Router Tool, achieving on-demand activation and efficient resource utilization.
• Long Straw Long Context ExtensionBased on our self-developed LongStraw technology, we have expanded the native 1M context window of GLM 5.2 to 2M. The core mechanism is to reuse shared prompts as a persistent state, avoiding the overhead of repeated calculations in long sequences, significantly reducing resource consumption and latency in long context inference, and enabling the model to remain efficient and stable when processing ultra-long documents and multi-turn complex dialogues.
• MinT Distributed Training PlatformMindLab's self-developed post-training infrastructure breaks through the bottleneck of traditional distributed training where Actors and Learners need to pass complete weights, instead only passing lightweight Adapters. This design is naturally compatible with the MoL architecture, supports the management of millions of Adapter catalogs, and can cover reinforcement learning training of trillion-parameter models end-to-end, significantly reducing communication bandwidth and storage pressure.
• Real-world driven reward functionUnlike training paradigms that rely on manual annotation or static benchmarks, Macaron emphasizes reward signals from real-world 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 validates the model's continuous understanding, information validation, and task replanning capabilities in a sandbox containing dynamic noise, dynamic environments, and dynamic users, ensuring that the model continues to evolve in real-world interactions.

How to use Macaron-V1

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

Macaron-V1's core advantages

• Capabilities are isolated and do not interfere with each other.The MoL architecture allows each expert to optimize independently, completely avoiding the performance fluctuations that occur when multiple capabilities are trained together.
• Extremely low cost expansionAdding new capabilities only requires attaching a new LoRA module; there is no need to retrain the entire model, allowing for flexible combinations.
• Long context supportThe LongStraw technique extends the context to 2M and reduces the overhead of long sequences through state reuse.
• Real-world evolution: We build our own dynamic evaluation benchmark, and the model continuously iterates in an environment that simulates real noise and dynamic users.
• Leading generation speedActual test result: 320 characters/second, far exceeding mainstream models on the market, with continuous and stable output after the first token response.

Macaron-V1 Comparison with Similar Products

Application scenarios of Macaron-V1