The real value in modern AI development is shifting from releasing finished model weights to sharing the training environments and reward systems that teach models to reason and solve problems. Xiaomi's recent release of its MiMo-V2.6 model demonstrates this trend, but the headline-grabbing benchmark scores mask a more significant story: the company is open-sourcing over 7,000 task environments, an end-to-end training framework, and lightweight agent tools that researchers can use to build their own reasoning models.

This move reflects a broader recognition in the AI research community that as model development increasingly relies on reinforcement learning with verifiable rewards (RLVR), the environments themselves become the critical bottleneck. RLVR is a training approach where models learn by attempting tasks whose outcomes can be automatically verified, like whether code passes a test or whether a software vulnerability was successfully reproduced. Unlike traditional reinforcement learning from human feedback (RLHF), which depends on subjective human judgment, RLVR can scale because the reward signal is objective and measurable.

What Makes Xiaomi's Release Different From Other Open-Source Models?

When Xiaomi livestreamed its reinforcement learning training process in mid-September, it exposed production metrics in real time over a five-day period, revealing that the RL phase alone cost approximately $854,000 for the smaller MiMo-V2.6-Flash model and $2,620,670 for the Pro version, totaling about $3.5 million. This level of transparency is rare; most AI companies keep training costs and processes confidential. The previous public benchmarks came from MiniMax, which disclosed that its RL phase cost $534,700, and DeepSeek, which reported $294,000 for its R1 model's RL training.

However, the cost disclosure is only part of the story. The real distinction lies in what Xiaomi is committing to release alongside the finished models. The company has released the model weights under the permissive MIT license, which carries no restrictions for commercial users, alongside its technical report and a 9-billion-parameter Qwen-based model designed as a starting point for further agentic reinforcement learning research.

"Releasing many high quality open-source RL environments is the most impactful thing anyone can do to push the open-source frontier right now. The equivalent of sharing high quality pretraining data, but in the new RLVR paradigm," said Thomas Wolf, co-founder and chief science officer at Hugging Face.

Thomas Wolf, Co-founder and Chief Science Officer, Hugging Face

Wolf's observation captures why the research community is paying close attention. In the era of RLVR-based model development, the environments in which models learn to solve problems are as valuable as the pretraining data that powered earlier generations of AI systems. Without access to diverse, high-quality task environments spanning software engineering, vulnerability reproduction, knowledge work, and web development, independent researchers cannot effectively train their own reasoning models.

How Does RLVR Differ From Earlier Training Approaches?

To understand why Xiaomi's environment release matters, it helps to understand the evolution of how AI models are trained after their initial pretraining phase. The field has moved through distinct paradigms, each with different requirements and constraints:

• RLHF (Reinforcement Learning from Human Feedback): This approach, which powered early versions of ChatGPT and other conversational AI, trains models to produce text that human raters prefer. It optimizes for subjective qualities like helpfulness and tone, but it tends to push models toward overconfidence because human raters often prefer confident-sounding answers.
• RLVR (Reinforcement Learning with Verifiable Rewards): This newer approach trains models on tasks with objectively measurable outcomes, such as whether code compiles and passes tests or whether a mathematical proof is correct. It created the reasoning models that excel at mathematics and coding but are slower and more computationally expensive to run.
• RLCD (Reinforcement Learning for Calibrated Decisions): An emerging third approach, exemplified by TypeSafe AI's Jev model, trains systems to return structured decisions with honest probability estimates rather than generated text. When the model assigns a probability of 0.8 to an outcome, that outcome should occur approximately 80 percent of the time across similar cases.

Each paradigm answers a different question about what to reward during training. RLHF rewards approval from human raters. RLVR rewards correctness on verifiable tasks. RLCD rewards calibrated probability estimates over bounded answer spaces. As the field has progressed from RLHF to RLVR and beyond, the infrastructure required to train models has shifted accordingly.

Why Are Researchers Calling This an "Impressive Level of Openness"?

Xiaomi's approach stands out because it blurs the distinction between "open-weight" and "open-source," terms that are often used interchangeably but mean very different things in practice. Many models released as "open" consist primarily of downloadable weights, the numerical values a model learned during training, while much of the process that produced them remains proprietary and inaccessible.

Some major AI companies have stretched the definition further. Meta, for example, has marketed its Llama models as open-source despite significant restrictions that have prompted open-source advocates to challenge that characterization. In contrast, Xiaomi's MIT license carries no conditions restricting commercial use, and the promised release of over 7,000 task environments, a complete training framework covering environment interaction through reward evaluation and policy optimization, and lightweight agent harnesses for experimenting with different tools and prompts represents a substantially deeper commitment to openness.

"Nearly half a year of silence. We spent it studying one problem: how far RL can scale. Three things we scaled: compute, environments and harnesses," said Fuli Luo, who leads Xiaomi's MiMo team after previously working at DeepSeek.

Fuli Luo, Lead, MiMo Team, Xiaomi

Luo's statement underscores that Xiaomi's team invested six months exploring how to push reinforcement learning further by increasing computational resources, diversifying the environments and agent setups, and scaling the resources dedicated to evaluating the model's attempts. The team committed to open-sourcing these various elements over the coming weeks following the model's release.

How to Leverage Open RL Environments for Independent Research?

For researchers and developers working outside major AI labs, access to production-quality RL environments and training frameworks opens new possibilities. Here are the practical implications of Xiaomi's release:

• Reduced Barrier to Entry: Independent researchers can now access the same types of task environments and training infrastructure that powered a frontier-level model, rather than having to build these from scratch or reverse-engineer them from published papers.
• Faster Iteration on Reasoning Models: With 7,000+ pre-built environments spanning software engineering, vulnerability reproduction, knowledge work, and web development, teams can focus on model architecture and training strategy rather than spending months constructing task harnesses.
• Standardization of RL Benchmarks: When multiple research groups use the same environments and evaluation frameworks, their results become more directly comparable, reducing the noise that comes from each team using proprietary or ad-hoc evaluation methods.
• Reproducibility and Verification: Open environments allow other researchers to verify Xiaomi's claims about model performance and to understand exactly what tasks the model was trained on, addressing a persistent criticism that much AI research lacks sufficient transparency.

Elie Bakouch, a former Hugging Face researcher now at Prime Intellect, highlighted the significance of this release on social media, noting that Xiaomi shipped the model and technical report less than one week after completing the final RL run, while also committing to release approximately 7,000 RL training environments and the framework that led to a top-ranked model on Artificial Analysis, an independent AI benchmarking service.

What Does This Shift Mean for the Future of Open-Source AI?

Xiaomi's move signals a recognition that the bottleneck in open-source AI development has moved. In the era of large language models trained primarily with RLHF, the main constraint was access to model weights and pretraining data. Now that the field is transitioning to RLVR-based reasoning models, the constraint has shifted to access to high-quality task environments and reward evaluation systems.

This transition has practical implications for how AI research gets distributed and built upon. Pretraining data, once released, remains relatively static. But task environments and reward systems are continuously refined and expanded as researchers discover edge cases and improve evaluation methods. By committing to open-source these components, Xiaomi is enabling a collaborative ecosystem where improvements can accumulate across many research groups rather than remaining siloed within individual organizations.

The MiMo-V2.6 model itself achieved strong benchmark results, with Artificial Analysis giving the Pro version an Intelligence Index score of 46, ranking it first among 114 large open-weight models tracked by that service. The flagship model is a trillion-parameter system with 42 billion parameters active at any given time, a one-million-token context window, and support for text, images, audio, and video. However, the model's performance numbers are secondary to the broader infrastructure Xiaomi is releasing, which may prove more valuable to the research community over time.

" }