Mind Lab, a Chinese AI startup founded in October 2025, says its Macaron-V1 model beats its own GLM-5.2 base on six of twelve benchmarks by training only four billion of the system’s 748 billion parameters. The company built the flagship version, called Venti, by freezing GLM-5.2’s 744 billion parameters and attaching four LoRA adapters of about one billion parameters each, one for chat, one for agentic tasks, one for coding, and one for interface generation. LoRA, short for low-rank adaptation, means training a small set of new weights bolted onto a frozen base model rather than retraining the whole system, a method that is far cheaper to train and swap than full retraining.
The architecture is what Mind Lab calls MoL, a mixture of LoRA adapters, and the system dynamically switches to whichever module best fits the task at hand. Andrew Chen, Mind Lab’s founder, has said models can share 99 percent of their parameters with their base while the remaining fraction determines how they differ. As users interact with Macaron, that usage data feeds a single LoRA adapter the company keeps retraining after every call, leaving the frozen base untouched.
That last mechanism is the company’s most interesting claim, and the one most worth interrogating. What actually updates is the adapter, a layer worth a few billion parameters at most. The 744 billion parameters of the frozen GLM-5.2 base do not change; they remain exactly as they were once pretraining finished. That is a real improvement over prompt engineering or retrieval-based memory, but it is not continual learning in the sense researchers such as Richard Sutton use the term, describing a system that learns from every experience without forgetting. It is closer to a narrow, swappable patch that specializes the model for a specific use case. The obvious risk: an adapter trained on the model’s own outputs reinforces whatever the model already tends to produce, drifting from ground truth with nothing but its own prior behavior as a check.
Mind Lab’s earlier Macaron-V1-Preview, released in June and built by attaching five roughly one-billion-parameter LoRA modules to GLM-5.1, is the version the company says outperformed GPT-5.4 and Claude Opus 4.6. Those results, like the six-of-twelve claim for the full Macaron-V1 released and open-sourced on July 21, come from benchmarks the company ran and published itself. KrASIA’s report does not cite independent verification of either result, and a derivative model built on GLM-5.2 beating GLM-5.2 is exactly the kind of specific, checkable claim that deserves a benchmark run by someone other than the company making it.
Mind Lab’s infrastructure numbers are more checkable in principle. In December 2025 the company ran reinforcement learning with LoRA adapters, a method it calls LoRA-RL, on Kimi K2, a trillion-parameter model, using 64 Nvidia H800 GPUs and roughly 10 percent of the compute a full-parameter reinforcement learning run would need. Mind Lab says no other team in China has made LoRA-RL work at that scale. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has published comparable results in the West. Mind Lab’s MinT platform, launched in January, now manages more than one million LoRA adapters and loads them roughly ten times faster than earlier versions, according to the company.
Macaron-V1-Preview reached $10 million in annualized revenue within two weeks of commercialization. Mind Lab’s parent company, Mindverse, has since raised $60 million in total funding, and Meituan’s investment arm led its nearly $50 million Series A, which closed in early 2026. The full Macaron-V1, including a smaller 50-billion-parameter Tall model built on Qwen 3.6 for local deployment, is open source. Both Venti and Tall support two-million-token context windows.
For a practitioner weighing adapter-based post-training against retraining a model from scratch, Mind Lab’s compute figures, roughly a tenth of the GPU cost of full-parameter reinforcement learning, are worth testing directly on an internal workload. Before accepting that Macaron-V1 actually beats GLM-5.2, wait for a benchmark run by someone outside Mind Lab, and ask exactly what data feeds the self-updating adapter before putting a system that trains on its own output into production.
This account is based on KrASIA’s August 3, 2026 report by Cheng Zi on Mind Lab’s Macaron-V1 and its LoRA-based continual learning approach.
