China's AREX-2, Built on a 27B Base, Beats a 1.6T-Parameter Model

A 27-billion-parameter model has beaten a 1.6-trillion-parameter system, according to Chinese researchers. The size gap is almost 60x. The AREX group from the Beijing Academy of Artificial Intelligence (BAAI), the University of Science and Technology of China (USTC) and Renmin University of China has presented AREX-2. The model is built on the lightweight Qwen3.8-27B base and trained on "long reflection" data drawn from machine learning and algorithmic programming tasks.

Reflection Instead of Extra Parameters

The usual route to a strong coding model is to add parameters, expand the corpus and apply reinforcement learning on verifiable tasks. AREX-2 takes a different path. The authors built their training data around a long "attempt, analysis, correction" loop. In that loop, the model does not stop after a single pass. It returns to its own solution, finds the weak spot and reworks the algorithm.

The domain was not chosen at random. In ML engineering and algorithmic programming, results are measurable: code either passes the tests or it does not, and a metric either rises or falls. That makes the feedback reliable, and long improvement trajectories can be collected and verified.

Why the 27B vs. 1.6T Comparison Draws Attention

The study's headline result rests on a contrast in size. If independent measurements confirm the claim, it will reinforce a long-running trend: on narrow, easily verified tasks, quality depends more and more on how the data is structured and how training is done, rather than on parameter count.

There is a practical angle for engineers. A 27B-scale model fits on a handful of accelerators and runs noticeably cheaper than trillion-parameter systems. It can be deployed in-house, fine-tuned and run in long agentic loops, where every extra token costs money.

RSI as a Working Framework

The authors tie the work to recursive self-improvement (RSI): a system's ability to iteratively improve its own algorithms. The idea is that a model that can refine code for ML tasks over long, meaningful iterations gradually gets closer to improving ML processes themselves. In their words, this opens the door to the autonomous evolution of intelligence. For now, this is a first step in a limited domain, not a fully self-improving system.

Key release parameters:

• Model: AREX-2
• Base architecture: Qwen3.8-27B (about 27B parameters)
• Training data: long reflection chains on ML and algorithmic programming
• Claimed comparison: outperforms a model at the 1.6T-parameter scale
• Developers: the AREX group; BAAI, USTC and Renmin University of China
• Research framework: recursive self-improvement (RSI)

What Teams Building Agents Should Check

Teams building coding and research agents should take a close look at the approach itself. If you have tasks with automatic verification, you can collect not only the final correct solutions but also the correction trajectories: what was tried, why it failed, how the approach changed. A dataset like this teaches a model to work with errors, not just to reproduce a finished answer.

Independent measurements are needed before adoption. First, teams should find out which benchmarks and under what conditions AREX-2 beats the larger model: the narrowness of the domain may heavily influence the conclusions. Then they need to confirm that the gain holds on tasks outside the training distribution and over long horizons, where errors accumulate. If those checks pass, compact models trained on reflection will become a real alternative to expensive giants for engineering agents.