MindLab launches Macaron-V1, a 748B LoRA-MoE agent model aimed at long-horizon tasks
智东西 · wechat · 2026-07-23
MindLab’s Macaron-V1 arrives as a MoE agent model family built around Mixture-of-LoRA, with two releases: the 748B Venti and the 50B Tall for local deployment.
Key points:
- The series is designed for long-horizon agent tasks, with benchmarks for chat, coding, agents, and GenUI, plus new internal tests like ChatBench and MacaronLivingBench.
- The article claims Venti is broadly competitive with top-tier models on public agent benchmarks, while remaining especially strong on long-term autonomous tasks.
- MindLab says the LoRA-heavy design cuts the cost of RL and adaptation to about one-tenth of full-parameter training in large sparse MoE systems.
- All variants are open-weight on Hugging Face.
- The company also says it began commercialization in July and reached $10 million ARR in two weeks, while raising nearly $50 million in June.
The piece argues that the next wave of AI will be less about one universal model and more about a shared base model plus many evolving LoRA adapters that preserve memory, preference, and long-term identity.
Related event: MindLab Open-Sources Macaron-V1, Betting on Continual Learning(4 posts)→
More from Companies & People
- Law Professor on Legal Engineering Jobs: Stigma Is Real but Builder Skills Open New Doors — jkubicki · 2026-09-11
- AI safety community mocked as 'bridge engineers' who say bridges can never be safe — Dan_Jeffries1 · 2026-09-11
- SoftBank's Masayoshi Son predicts 100 trillion self-replicating AIs: "humans' era as top life form is ending" — Puzzleheaded-King584 · 2026-09-11
- IIT Madras Launches EdTech Tulna Standards for AI-Powered Learning Products — ravi_iitm · 2026-09-11
- Warp's six non-engineering teams all run on Linear and Claude Code — mon__lim · 2026-09-11
- Anthropic Insiders: Not Everyone at the Lab Believes in High p(doom) — anpaure · 2026-09-11