"Mindverse" Closes Pre-A Series Led by Ant Group, with HSG Among Co-Investors | Linear Portfolio

Mindverse (maker of Macaron AI, with research lab Mind Lab) has recently completed a Series Pre-A round led by Ant Group, with HSG and others co-investing, raising over $20 million in total.

Mindverse wants to use LoRA to give everyone their own large model, achieving true "Personal Intelligence" at scale. Its experimental product Macaron AI launched in August 2025 and quickly gained market traction. Since then, the team has completed multiple 10x iterations, with key metrics demonstrating its ability to turn research into product.

Linear Capital was Mindverse's angel-round investor. We're thrilled to see the team's execution and growth since the early days of midreal.ai.

---

On January 27, Personal Intelligence company Mindverse (maker of Macaron AI, with research lab Mind Lab) announced the completion of its Series Pre-A round led by Ant Group, with HSG and others co-investing, raising over $20 million in total.

This round marks a turning point: Personal Intelligence — individually owned models — is moving from concept to scalable reality.

For the past two years, most products have tried to serve as many users and scenarios as possible with a single, more powerful general-purpose model. Today we can see the limits of this approach. One model for everyone means the model must be "averaged out" — it can be smart, but it can never truly know you. Even products that introduce multi-model routing are still essentially assembling general capabilities, rather than enabling a model to develop a continuously evolving individual personality and long-term memory.

What users actually want from "thousand faces for thousand people" isn't more complex prompt templates or larger external knowledge bases. It's a model that remembers preferences, absorbs feedback, and keeps updating itself — becoming more like your personal assistant over time. Mindverse's core bet is that as the returns from scaling general model parameters diminish, the next phase of scaling won't just be "bigger" — it will be giving every user their own model.

Mindverse aims to use LoRA to solve the key challenge of making large models truly personalized. In Mindverse's architecture, LoRA serves as a continuously trainable parameter layer that carries long-term memory and personality traits, allowing the model to keep learning from user feedback after pre-training.

The Mindverse team was founded in October 2023. Early on, they co-authored the paper FireAct with Shunyu Yao, starting from agent fine-tuning research and gradually evolving to LoRA supervised fine-tuning. By August 2025, they achieved the first-ever large-scale LoRA reinforcement learning system at massive parameter scale.

Based on team disclosures and public information, only a handful of teams globally have achieved trillion-parameter LoRA reinforcement learning — among them, Mindverse and Thinking Machines Lab, founded by OpenAI's former CTO.

To productize and platformize this capability, Mindverse has built the trillion-parameter training platform Mind Lab Toolkit, now open for external companies to apply for testing.

Under this vision, Mindverse's experimental product Macaron AI launched in August 2025 and quickly attracted widespread market attention. Macaron AI is a Personal Intelligence product for personal life scenarios, designed to let users experience the real value of having their "own model."

Since launch, the team has completed multiple 10x iterations, with key metrics showing its ability to turn research into product:

• App generation time: from roughly 20 minutes at launch, optimized to ~2 minutes through RL training, and further reduced to 10-second range in recent versions;
• Cold start time: using a self-built container solution to support large-scale individualized online serving, cut from ~40 seconds to near 2 seconds;
• Success rate: improved from ~90% early on to 99.9%;
• User-created mini-apps on the platform: surpassed 300,000, creating sustainable scenario and demand density.

According to the team, the product currently runs 28 different LoRA models based on user type. More critically, Mindverse is breaking through to larger-scale deployment capabilities, with a target of Q2 2026 to achieve one customized model per user for hundreds of thousands of users — truly pushing "thousand faces for thousand people" to "one model per person."

Founder Kaijie Chen said: "The hyper-personalized experiences that models can enable are just beginning to emerge. We want to give everyone their own model, where every person's preferences are remembered — that's the future of AI."

Mindverse's ambition goes beyond building a better app. It's driving a larger shift: evolving AI from a universal tool for everyone into a personal intelligence for each individual. After the first half of the general-model era has been proven out, Mindverse is betting on — and building — the second half.