MindLab open-sources Macaron-V1, a GLM-5.2 family built on continual learning
机器之心 · wechat · 2026-07-24
MindLab has open-sourced Macaron-V1, a GLM-5.2-based model family built around continual learning and multi-LoRA collaboration.
- The flagship Venti variant combines a frozen 744B GLM-5.2 base with four 1B LoRA experts for chat, agent work, coding, and generative UI.
- MindLab argues that post-training should not end when parameters freeze: the model should keep learning from deployment-time experience.
- Its Mixture-of-LoRA design treats LoRA adapters as persistent, writable local state rather than just cheap fine-tuning patches.
- The article cites scaling results showing LoRA RL can be cheaper than full-parameter RL while still improving performance on large MoE models.
- On the systems side, MindLab pairs the model with MinT, MindForge, HCP, and LongStraw to align training, serving, long-context RL, and agent harnesses.
- The release also includes open weights for Macaron-V1-Venti, Macaron-V1-Coding-Venti, and Macaron-V1-Tall on Hugging Face.
Related event: MindLab Open-Sources Macaron-V1, Betting on Continual Learning(4 posts)→
More from coding & agent
- Astra storyboards plus Minimax H3 per-shot generation boost video success rates — Hailuo_AI · 2026-09-11
- Codex tip: use Sol with Astra and Luna sub-agents to save usage — pvncher · 2026-09-11
- agents-best-practices: a provider-neutral Agent Skill for designing and auditing agentic harnesses — tom_doerr · 2026-09-11
- Cognition's SWE-2 uses a KKT duality argument in RL to shift the effort Pareto curve — YouJiacheng · 2026-09-11
- First-ever Three.js Conference lands in Paris, with a panel on AI-shortened design workflows — OdinLovis · 2026-09-11
- Agile co-author Ron Jeffries publishes 'Resist AI', urging developers to push back — mborch · 2026-09-11