MindLab launches Macaron-V1 and says continual learning is the next AI frontier

新智元 · wechat · 2026-07-24

MindLab unveils Macaron-V1 and claims a fast path to continual learning

The article compares a wave of AI labs betting on LoRA-based post-training and continuous learning. It highlights Mira Murati’s Thinking Machines Lab, Richard Sutton’s new OakLab, and DeepSeek founder Liang Wenfeng’s view that the next major step after agents is continual learning.

What MindLab released

Macaron-V1 comes in two versions:
Venti: a 748B-parameter flagship built on GLM-5.2, with four 1B LoRAs and a 2M-context window.
Tall: a 35B version based on Qwen-3.7 that can run on a Mac.
The architecture separates skills into four LoRA modules for chat, agents, coding, and GenUI, with a router switching between them at runtime.
The team calls this Mixture-of-LoRA (MoL): shared base model, independent specialized adapters.

Claimed results and engineering details

On SWE-bench Verified, Macaron-V1-Venti reportedly scores 85.6, ahead of DeepSeek-V4-Pro, MiniMax-M3, and Kimi-K2.6.
On Deep-SWE, it reportedly reaches 58.4, versus 46.2 for its GLM-5.2 base.
MindLab says the model and its deployment harness were co-designed and trained together, so tool use and long-horizon workflows match production behavior.
A REPL-style harness lets the system save intermediate state and reuse successful workflows instead of starting over each time.

Why the article treats it as important

MindLab says it ran LoRA RL at trillion-parameter scale, a setup the article describes as rare globally.
It also argues that learning should continue after deployment, with user feedback becoming persistent model state rather than being lost after inference.
The piece frames LoRA not as a cheap substitute, but as a way to let models accumulate durable, reversible, specialized experience.

Business angle

MindLab reportedly raised $50M in a Series A led by Meituan, with Ant Group, Sequoia China, and ZhenFund also participating.
The lab says it opened API commercialization in early July and reached $10M ARR in two weeks, which the article presents as unusually fast for a model company.
It says customers are buying the ability for the model to keep learning inside their own workflows, not just raw tokens.

Broader thesis

The article argues that the AI race is shifting away from “who trains the biggest base model” toward “who can make deployed models keep getting smarter.” It presents MindLab as an especially efficient player because it builds on Chinese open-source base models instead of training everything from scratch.

Original post →

MindLab launches Macaron-V1 and says continual learning is the next AI frontier

More from Companies & People