What Does Open Source Mean for the Lucrative RL Environment Business?

China’s generosity to the open-source AI ecosystem has delivered yet again. And just when Xiaomi, the consumer tech giant, seemed to have surprised the world enough by chasing the frontier with its open-weight MiMo family of models, it has gone a step further. 

It has now open-sourced the reinforcement learning (RL) environments used to train its new model, MiMo-V2.6. 

The repository, available on Hugging Face, consists of 7,780 tasks across five domains: coding, cybersecurity, knowledge work, web development, and music. These give AI agents environments in which they can perform tasks, receive feedback and learn from the outcomes. 

RL has been a driving force behind the era of reasoning AI models. This method trains models through trial and error, rewarding them when they arrive at the desired outcome and penalising them when they do not.

Over repeated attempts, the model learns which actions and reasoning patterns are more likely to produce better results.

RL environments provide the setting in which this process takes place. They give the model a task, the tools to act on it, and a way to evaluate the outcome. 

In a coding environment, for example, an AI model can modify a codebase, run tests and receive a reward when the tests pass. 

The environment effectively becomes the world in which the model practises, while the feedback it provides becomes the signal the model learns from.

Another example from MiMo’s RL environments is one that recreates a financial-diligence workflow around Argo Group, the specialty insurer that was acquired by Brookfield Reinsurance for $1.1 billion in 2023. 

The agent takes on the role of a financial-diligence analyst preparing a year-end reinsurance review and is given access to simulated business systems and source files, including SAP Reinsurance Administration, a reinsurance recoverables CSV and a financial model. 

It must reconcile the company’s recoverables, identify how specific legal entities roll up, compare figures across different records and explain discrepancies.

Worth Millions?

In the race to deploy agentic AI workflows and frontier AI models into production, RL environments hold immense value. 

“To put [MiMo’s RL environments] in perspective—if you have to buy some tasks like this, it’s usually hundred to thousand dollars per task, so this repository is literally worth millions,” Harveen Singh Chadda, an LLM researcher at Sarvam, wrote on X. 

An estimate from EpochAI earlier this year revealed the economics of RL environments. 

It found that individual RL tasks typically cost $200 to $2,000 each, while particularly complex software-engineering tasks can reach around $20,000.

Environment contracts can run into the hundreds of thousands or even millions of dollars per quarter. Epoch also noted that one estimate put the compute cost of running RL training on a single task at around $2,400. 

The demand has also opened up an ecosystem of startups building RL environments for specific production workflows. 

Companies such as Mercor, Surge, Handshake, and Turing, which traditionally built their businesses around human-labelled data, now also sell RL environments.

Newer companies have emerged with RL environments at the centre of their businesses.

Deeptune, which built high-fidelity simulations of workplace software such as Slack and Salesforce, raised $43 million in a Series A led by Andreessen Horowitz in March, before being acquired by Mercor three months later. 

Mechanize, which builds environments and evaluations for frontier coding agents, raised $9.1 million in April.

AfterQuery, which builds expert data and agent environments around APIs, MCP servers and developer tools, raised $30 million in April and was subsequently valued at $3.2 billion in a 2026 funding round.

So, do MiMo’s RL environments put a dent in this ecosystem? Does it ease the lives of AI engineers wanting to perform RL runs on their models? 

How Will It Affect RL Environment Builders?

“This [MiMo RL environments] shows how frontier teams are thinking about RL and how they are structuring it,” says Sachin Dharashivkar, an RL researcher and founder of Savante AI, to AIM.

 “Chinese AI teams have spent significant time experimenting with things like these, and they are close to the frontier models.” 

While developers are expected to explore training models with these environments, Dharashivkar believes the release could deliver even better value by teaching people how to construct them.

“Because the fun part about AI models is that each new one needs to be different. As these RL environments have become open source, everybody will train on them. And when everybody trains on the same thing, those capabilities become widespread,” he says. 

And thus, he believes it will be of less help in building a differentiating factor for model companies, meaning companies that build RL environments will still hold value. 

“Anyone who is entering the domain of RL as a company—this is useful in the sense that it teaches them the best practices, and they don’t need to worry about whether their demand is going away because model companies now have access to 7,000 environments.”

He concludes that the release is, in fact, a boon for RL environment companies. Studying the repository used to train a near-frontier model like Xiaomi’s MiMo-V2.6 gives them a clearer sense of the quality and capabilities that model companies are likely to expect from RL environments.

What open source threatens most, therefore, may be the value of generic, reusable environments rather than the entire market. 

Epoch AI found that exclusive environments can command roughly four to five times the price of non-exclusive ones, suggesting that proprietary access remains an important part of the business.

Besides, MiMo’s release may commoditise access to a large collection of RL environments, but it does not commoditise the infrastructure required to run them at scale.

For example, one startup called Prime Intellect makes its Environments Hub open source, while building a commercial stack around it that includes hosted training, evaluations and secure sandboxes capable of running thousands of environments concurrently.