Regarding the agenda for the next phase of large language models, we have found a Chinese Neo Lab that specializes in continuous learning.
By Wang Xinyi
Edited by Zhang Yuxin
I met Chen Kaijie, founder of Mind Lab, at 9:30 p.m. in Beijing, and he had been meeting investors for a whole day.
Chen Kaijie is a serial entrepreneur. He suspended his studies at Duke University, launched the AI interactive story platform MidReal, and also rolled out the Personal Agent application Macaron, which topped the Product Hunt daily chart on the day of its launch. In October 2025, Mind Lab was founded with a team of about 30 people. Andrew Chen, founder of Mind Lab, once co-authored the FireAct paper with Yao Shuny u. The team members mainly come from enterprises and universities such as xAI, DeepMind, DeepSeek, ByteDance Seed, MIT and Tsinghua University. They reappeared in the public eye under the identity of Neo Lab.
The direction they are betting on is highly consistent with the lifelong learning direction emphasized by Turing Award winner, father of reinforcement learning Richard Sutton and former CTO of OpenAI Mira Murati.
A few days ago, at the main forum on the first day of WAIC 2026, Sutton delivered a speech emphasizing that the core route of the next generation of AI is experience-driven, and the static labeled data mode has reached its ceiling. DeepSeek has also recently brought lifelong learning into the vision of more people, claiming that the problem to be solved after Agent is lifelong learning, which is also a necessary capability for the next generation of models.
It can be seen that post-training and lifelong learning are increasingly becoming the key criteria for evaluating model capabilities in the next stage.
Mind Lab released the Macaron-V1-Preview model last month, and its business has been pushed forward rapidly since then: just two weeks after the commercialization was launched, its ARR reached 10 million US dollars.
The composition of Macaron-V1-Preview is as follows: 5 LoRA expert modules with parameters of about 1B are mounted on GLM-5.1. Although it is only a preview version of the model, Macaron-V1-Preview has performed brilliantly on multiple Benchmarks. It not only outperforms the base model GLM-5.1, but also leads a number of models such as GPT 5.4 and Claude Opus 4.6. The capability beyond the base model comes from the LoRA mounted on it.
What really makes them attract wide attention is the post-training path using MoL (Mixture-of-LoRA). When the model performs different tasks, the system can dynamically switch to the most suitable expert module according to the task type. In the long-term use of users, the data accumulated can also precipitate an exclusive LoRA, which can be continuously updated with the continuous calling of the model.
The Preview version has first verified the technical path and completed market validation. Shortly afterwards, on July 21, Mind Lab released and open-sourced the official version of Macaron-V1. According to the officially released benchmark, Macaron-V1 won 6 SOTAs in 12 benchmark tests, and the rest of the results are also close to the cutting-edge models.
This version has two models: the flagship version Venti is a model with a total parameter of 748B that undergoes post-training based on GLM-5.2, of which 744B comes from the frozen GLM-5.2 base and 4B comes from 4 groups of LoRA trained by Mind Lab itself, each with about 1B parameters, which are responsible for four capabilities: Chat, Agent, Coding and UI generation. The other is the lightweight version Tall for local deployment, which is a model with a total parameter of 50B that undergoes post-training based on Qwen3.6. Both versions natively support 2 million token context.
This also means that by only changing 4B parameters, this team enables GLM-5.2 to achieve capabilities beyond its original performance.
For Mind Lab, entrepreneurship is a process of constantly maintaining the main line and overturning and redoing work. Stripping off all the external noise, a more far-reaching technical route within the team is evolving continuously — lifelong learning.
At the same time, leading capitals have also placed bets one after another. Since its establishment, Mindverse, the parent company of Mind Lab, has raised a total of 60 million US dollars in financing. At the beginning of 2026, it completed a nearly 50 million US dollars Series A financing led by Meituan Strategic Investment, with additional follow-on investment from Yuanhe Puhua, Shokz Technology, Bianyuan Capital and existing shareholders. Historical shareholders include Ant Group, HSG, Yuanyi, ZhenFund, Gaorong and others.
Intelligence of Experience: Making the Model Understand You Better
The original starting point can be traced back to a FireAct paper co-authored by Andrew, co-founder of Mind Lab, and Yao Shuny u in 2023. At that time, they believed that to improve the task performance of Agent, instead of doing prompt engineering, it is better to directly train the relevant data into the model.
This is the starting point for them to bet on lifelong learning and post-training. At that time, this was not yet a widely concerned technical route. "When we started doing lifelong learning, we didn't even know it was called lifelong learning," Chen Kaijie said.
The problem with existing technologies is that a general large model is difficult to adapt to all scenarios. After the model training is completed, its parameters are usually fixed, and the model parameters will not change according to specific scenarios, while the cost of training the model from scratch is very high.
If you want a large-parameter model to perform tasks in different vertical scenarios, a complex Harness can certainly achieve good adaptation to the scenarios, but the inevitable consequence at the same time is that it will consume a lot of Tokens and run very slowly, which obviously cannot solve the problem fundamentally.
Under the framework of Shannon's information theory, in the reinforcement learning paradigm, the demand for the number of parameters learned by the representation model becomes smaller, which also means that under the condition that the model parameters are large enough and sparse enough, the effect of LoRA in reinforcement learning in the same vertical field can reach the effect of full-parameter training.
Therefore, there is no need for full-parameter retraining to achieve good adaptation to specific scenarios.
The first achievement they delivered is to realize reinforcement learning with trillion-level parameters. In December 2025, Mind Lab implemented end-to-end LoRA reinforcement learning training on Kimi K2 (a trillion-parameter MoE model). With only 64 H800 GPUs, it achieved an effect close to full-parameter training, and the GPU consumption was only about 10% of that of traditional full-parameter RL.
At that time, some large factories and startups such as ByteDance, Alibaba, DeepSeek and Kimi also had the capability of trillion-parameter reinforcement learning. However, Mind Lab was the only team in China that could run LoRA-RL on trillion-parameter models, and the only team overseas that could achieve this was Thinking Machines Lab founded by former OpenAI CTO Mira Murati.
The difficulty of realizing trillion-parameter reinforcement learning lies in that reinforcement learning has very high precision requirements for the infrastructure. If the precision of training and inference is inconsistent, deviation drift may occur, leading to the failure of the final result to converge. The solution of Mind Lab is to design a hybrid collaborative parallel engine that integrates tensor parallelism, pipeline parallelism, expert parallelism and sequence parallelism, enabling LoRA to run stably on the MoE architecture. They also solved the mismatch problem between training and inference, and introduced truncated importance sampling to correct the distribution difference.
The delivery of this result was a brilliant victory for Mind Lab. However, in Chen Kaijie's view, the real technical moat of the team is the infrastructure.
A year ago, the data volume of post-training was about one tenth of that of pre-training, but now, the post-training data volume in many models is even larger than the pre-training data volume. Building a good post-training infrastructure is more difficult than pre-training.
Therefore, in January this year, they launched an infrastructure platform for LoRA training and inference — MinT, which can provide customers with full-process solutions for large model post-training. Customers can obtain computing power support on this platform, and can also train their own LoRA. MinT can manage millions of LoRA models, and only transmits the extremely lightweight LoRA Adapter during the process of training, evaluation, deployment and rollback, with the real-time loading speed increased by nearly ten times.
Subsequently, they devoted more energy to the route of model self-evolution and lifelong learning, and continued to explore the possibility of MoL (Mixture of LoRA, LoRA collaboration) on the basis of LoRA.
Lifelong learning is gradually evolving from a niche exploration of a few people to industry consensus. Two weeks ago, Sutton, the father of reinforcement learning, started his own business and founded the new company Oak Lab, committed to building an agent that can continuously learn from its own experience and evolve in real time. Tang Jie, founder of Zhipu AI, also mentioned in a recent internal letter that the key technologies to be conquered in the next step include memory, lifelong learning and self-evaluation. Leading teams at home and abroad have entered this track one after another, and the acceleration of this track is visibly increasing.
In Chen Kaijie's view, at the current stage, there are four solutions for lifelong learning:
The first is the context of chat, where the model repeatedly understands the user's intention within the conversation window;
The second is External Memory, which uses RAG as a memory extension. As more and more things need to be remembered, the database also expands at a very fast speed, and the problem-solving capability of the model is directly linked to the database retrieval capability;
The third is Harness and Loop Engineering. The problem is that as the complexity of Harness increases, it may consume a lot of Tokens and run very slowly;
The fourth is to directly modify the parameters of the model to adapt to the application scenario, turning it into a vertical model from the bottom layer, so as to fundamentally change the performance of the model.
LoRA belongs to the fourth type of solution. "After we confirm that additional learning is required for a certain scenario, we will enable the LoRA mode," Chen Kaijie told us.
General large models cannot solve 100% of the problems in specific scenarios. To achieve full adaptation, it is necessary to carry out vertical training of the model in any scenario. Sometimes 99% of the parameters between different models can be shared, but that 1% of the part determines their differences.
Based on this understanding, Chen Kaijie believes that models should be put into different experiences to allow them to grow continuously.
Mind Lab summarizes this set of ideas as "Experiential Intelligence": the model can learn from its own experiences, convert experiences into new capabilities, which enable it to solve more difficult problems, and more difficult problems bring richer experiences, so that intelligence becomes a continuously growing process. Moreover, this process is automated, and the model needs to continuously self-evolve and self-iterate to achieve real "lifelong learning". This direction also coincides with Sutton's "experience-driven" route.
The successive launch and market validation of Macaron-V1-Preview and Macaron-V1 have also made the public see the possibility of experiential intelligence.
Intelligence of Collaboration: Breaking the Upper Limit of Intelligence with Mixture of LoRA
In Chen Kaijie's view, post-training is gradually becoming an independent track. At present, a number of startups focusing on undertaking pre-trained models and focusing on post-training have emerged in the industry. Among them, many players focus on small model scenarios, doing post-training and data processing related work in vertical segments.
Mind Lab is not fighting alone on the LoRA route. Chen Kaijie told us that their technology is highly consistent with that of Thinking Machines Lab, especially in choosing the LoRA training route. TML also independently confirmed the same conclusion in its paper "LoRA Without Regret": Using LoRA for reinforcement learning on a sufficiently large MoE model will not cause any performance loss.
However, reinforcement learning also needs to be coupled with the model architecture. Due to the different architectures, the models that Thinking Machines Lab and Mind Lab can train are somewhat different. For example, the GLM-5 series introduces efficient inference architectures such as MTP (Multi-Token Prediction) and DSA (Dynamic Sparse Attention), and these designs have specific adaptation requirements for the training framework, while the entire technology stack of Thinking Machines Lab is built for the DeepSeek-V3 standard architecture, so it is difficult to be compatible with GLM.
Mind Lab took the lead in completing the reinforcement learning post-training on GLM-5.1 and GLM-5.2. This is the world's first external team achievement of completing reinforcement learning post-training on GLM-5.1.
It is worth mentioning that whether it is Macaron-V1-Preview or Macaron-V1, Mind Lab has set up multiple LoRAs, which can be used alone or in collaboration. Behind this setting is their discovery that a single LoRA cannot comprehensively improve all capabilities of the model, but using multiple LoRAs on one model can greatly improve the overall capability.
Mind Lab once used 200 different sets of data to train 200 LoRAs respectively, and let these LoRAs mount on a certain model to complete tasks in collaboration. They have two findings:
First, after the number of LoRAs involved in collaboration increases, the performance of model tasks will also show a log-linear growth relationship. The more LoRAs collaborate, the better the effect.
Second, compared with one LoRA trained with the 200 sets of data, the performance of the model after 200 LoRAs collaborate together is better, with an improvement of about 25% in capability compared with the former.
This also makes Mind Lab see the upper limit of intelligence that can be broken through by model collaboration. A single LoRA is not enough, and MoL (Mixture of LoRA) may be a more effective approach. "This is also a topic that needs continuous exploration, such as how many LoRAs in collaboration are more meaningful, how to divide the work between LoRAs, when collaboration is needed and when a single LoRA is sufficient."
The competition for post-training in China is also accelerating. Players in the track have their own strengths, but there are still not many teams that can run LoRA-RL at the trillion-parameter scale, build a million-level LoRA management infrastructure, and implement lifelong learning into commercial scenarios at the same time.
From Post-Training to Lifelong Learning
Mind Lab's current business is divided into three parts: continuous cutting-edge research on the Infra system and post-training; update and iteration of the C-end product Macaron; and good customer service on the B-end, on the one hand, deploying cheaper training Infra for enterprises such as Microsoft Cloud and Huawei Cloud, and on the other hand, providing MinT and lifelong learning LoRA models.
Among them, the Macaron App is positioned as a personal Agent for life scenarios. Users can generate exclusive small applications through natural language, and provide daily life assistance and emotional companionship based on long-term memory and daily conversations. As an application in C-end scenarios, the user interaction data accumulated by Macaron is also of great value for model training, and the behavior performance in these real usage scenarios can help the team understand the interaction mode between users and models more deeply.
It is worth mentioning that many of the B-end customers are smart hardware manufacturers, and smart hardware is exactly the natural scenario for the implementation of lifelong learning models. The data accumulated by different users in use is almost different, and the personalized and differentiated data is very valuable for the personalized lifelong learning of models in vertical fields.
On the basis of basic model calling services, Mind Lab will also provide customers with a "lifelong learning" option in the future. After checking this option, all kinds of high-performance data generated in the use of the model will be precipitated into an exclusive LoRA, and continuously train a model that is more in line with the customer's usage scenario and has higher output accuracy. In the process of lifelong learning, LoRA can almost update a new version every day to make it better adapt to user needs and scenarios.
However, for Mind Lab, commercialization is only the verification of technology implementation at the current stage, not something they are in a hurry to do. Kaijie told me that Mind Lab has reached 10 million US dollars in ARR, and there is still much room for growth. "But we don't want to use all the computing power to run orders, the focus still has to be on doing a good job in research."
For Mind Lab, the background of Research will always remain unchanged. Chen Kaijie told us that there are currently more than 30 members in the team doing research, including basic Infra and cutting-edge research.
"Promoting the progress of LoRA is the role we want to play. We need to solve the last few kilometers of work after the model pre-training is completed." LoRA is a technical direction with many variants, such as how to do linear attention, how to expand the context of the model, how to achieve good collaboration of LoRAs, etc., all of which are yet to be explored.
Image Source | Provided by the enterprise
