Conversations with investors lately point to a shift in artificial intelligence: attention is returning to large models and the infrastructure behind them.

As enthusiasm for applications cools, investors are looking more closely at the systems used to build and run models. Platforms for post-training and inference, the process of running a trained model, have drawn particular interest.

On September 22, Mind Lab unveiled Mint Recursive, a post-training and inference platform for industry use, and Macaron-V1.1, a model post-trained on the platform.

Macaron-V1.1 is a 752-billion-parameter model post-trained from GLM-5.3. In July, Mind Lab founder Chen Kaijie told 36Kr that less than a month separated Macaron-V1-Preview, post-trained from GLM-5.1, from Macaron-V1, based on GLM-5.2. The latest update took less than two weeks.

Macaron-V1.1 combines a 744-billion-parameter base model with four two-billion-parameter expert modules built using low-rank adaptation (LoRA). Each specializes in one of four areas: chat, agents, coding, and generation.

The LoRA modules have doubled in size from one billion parameters in V1 to two billion as Mind Lab tests the ideal size for its specialist modules. According to the company, V1.1 improved on V1 across all six agent benchmarks, with particularly strong performance on SWE-Marathon, which measures an AI model’s ability to handle extended coding tasks.

Macaron-V1.1 also serves as a demonstration of Mint Recursive: all of its post-training took place on the platform.

Mint Recursive covers data, evaluation, post-training, deployment, and inference. Users can access its training and deployment tools through application programming interfaces (APIs) and a Python software development kit (SDK). The platform can also collect feedback from models in use, allowing them to keep learning from real business tasks.

Building the infrastructure for post-training

Mind Lab’s work on models and infrastructure has developed together.

In January, the company introduced an early version of Mint Recursive, offering LoRA training and inference as part of a complete post-training service.

Other model developers are also investing in infrastructure.

In a post on X, Z.ai founder and chief scientist Tang Jie described the company’s efforts to use AI to improve AI. An infrastructure agent powered by GLM-5.3 automatically optimized the model’s own inference engine. Tang said it tripled end-to-end throughput from the initial baseline in less than two weeks on a cluster of more than 100,000 Chinese-made chips.

Z.ai is using AI to improve its inference infrastructure. Mind Lab is also involving models in their own optimization, but with a different objective.

It wants to improve the full workflow from post-training to deployment and inference, including the supporting data, evaluation, and feedback collection.

Mint Recursive has three main modules: training and deployment, an environment hub, and data and experience. It supports supervised fine-tuning (SFT), reinforcement learning (RL), direct preference optimization (DPO), full-parameter fine-tuning, and other training approaches. It also works with open-source models of various sizes from families including GLM, Qwen, DeepSeek, Kimi, and MiniMax.

Once training is complete, models can be deployed with one click. The evaluation system then helps prepare for the next round of training.

In September 2025, Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, published a technical blog titled “LoRA Without Regret.” Its analysis and experiments showed that LoRA, when configured appropriately, can match the performance of full-parameter fine-tuning.

LoRA is now a widely used approach to post-training.

Mind Lab has long focused on LoRA. Mint Recursive makes both training and deployment serverless: customers do not need to buy GPUs or manage computing clusters and instead pay based on token usage.

LoRA trains a relatively small set of additional parameters. On Mint Recursive, it allows models to keep learning from business feedback and improve at specific business tasks.

The aim is to help models understand a company’s business better while substantially reducing post-training costs.

For deployment, each of the four specialist modules operates as a lightweight adapter attached to the same base model.

The adapters do not interfere with one another and can be upgraded, launched, or taken offline independently. Earlier versions remain available for comparison and rollback. Customers can also merge an adapter into the base model to create a standalone model.

Sharing a base model allows companies to target post-training at individual business needs. This can reduce the cost of post-training, deployment, and updates while making model weights more flexible to adjust.

All of this depends on infrastructure for data, evaluation, and feedback. Building that infrastructure grew out of Mind Lab’s work on post-training models.

The company began laying the groundwork before Mint Recursive’s release. In 2025, Mind Lab began researching post-training for trillion-parameter models such as Kimi K2. It started building the infrastructure that would become Mint Recursive around that time.

By the end of 2025, it had contributed what it described as the industry’s first trillion-parameter LoRA-RL implementation to Nvidia Megatron-Bridge, gaining experience in lightweight reinforcement learning for very large base models.

Earlier this year, Mind Lab contributed LoRA and reinforcement learning solutions to several open-source projects, including Volcano Engine’s reinforcement learning training framework and Nvidia’s next-generation training backend, Megatron Lite.

In July, the Macaron-V1 series was officially released and open-sourced, validating the effectiveness of the training approach on tasks including agents and coding.

Not every company can, or needs to, build that infrastructure itself.

Post-training trillion-parameter models presents difficult engineering problems: allocating GPU memory across distributed training systems, maintaining numerical consistency between training and inference, and updating model weights while asynchronous reinforcement learning continues. Each can be a substantial challenge for a startup.

Chen told 36Kr that reinforcement learning is relatively easy for small models but becomes much harder as models grow. Mind Lab adds another challenge by combining it with continual learning based on LoRA.

Mint Recursive offers three ways to use its services: specialists can work alongside an enterprise on training; customers can run automated training through the platform; or they can train models themselves using APIs and the Python SDK.

Training a company’s own model

Demand for company-specific models is helping drive interest in lightweight post-training and continual learning.

Businesses may have the most pressing need for models tailored to their work.

“General-purpose large models can cover only 20–30% of use cases across industries. The remaining 70–80% require domain-specific continual learning for optimization,” Chen said.

Growth overseas points to demand for these services. Fireworks AI, a Silicon Valley provider of inference and post-training infrastructure for open-source models, reached a valuation of USD 17.5 billion in July 2026. Its annual recurring revenue surpassed USD 1 billion, while daily token volume exceeded 40 trillion.

In a discussion with Sequoia Capital, Fireworks AI CEO Lin Qiao said post-training is no longer a game reserved for model teams. Once AI products find product-market fit and begin scaling, post-training becomes difficult to avoid.

AI product companies have an advantage distinct from that of model providers: data from real user activity, including tasks, feedback, and preferences.

More than 95% of traffic on Fireworks’ platform comes from specialized models that customers have fine-tuned themselves, rather than general-purpose base models.

Qiao describes the idea as “owning your own intelligence.”

More companies are choosing to have application teams build models themselves. Competition is increasingly extending to the experience embedded in those models: a company’s judgment, taste, business data, and understanding of its customers.

Much of that knowledge comes from daily operations. An application company knows what users want to accomplish, what counts as a good result, where failures recur, and how professionals in its field judge the outcome.

Turning that experience into capabilities a company can retain in its own model remains a challenge.

Mint Recursive aims to help businesses develop their own evaluation frameworks, reward mechanisms, and training data so they can keep improving their models.

As a model serves users, the platform records the full sequence of steps in each interaction, whether it succeeds or fails.

Failed interactions can be particularly useful. Once the platform identifies what went wrong, why, and how to correct it, those judgments and records can become new training data.

Retraining the model on that data and deploying it again produces fresh feedback for the next round of training.

The environment hub supports this process.

According to a Nvidia technical blog, reinforcement learning for agents requires an environment to define the dataset, verifier, and state. The environment is not only where a model trains. It is also where it is evaluated, meaning evaluation, synthetic-data generation, and reinforcement learning can all take place within the same environment.

Mint Recursive’s self-improvement process is still at an early stage. Humans remain involved, and fully automated self-training is some distance away.

For customers, the starting point is the experience they already hold: knowledge of real business operations and the distinctive data those operations produce.

KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Wang Xinyi for 36Kr.