China's 27B AREX-2 Outperforms a 1.6 Trillion-Parameter Model

A model with 27 billion parameters has, according to Chinese researchers, outperformed a system with 1.6 trillion. The size gap is nearly 60-fold. 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 foundation and trained on "long reflection" data drawn from machine learning and algorithmic programming tasks.

Reflection instead of extra parameters

The usual path to a strong coding model is to scale up parameters, expand the corpus, and add reinforcement learning on verifiable tasks. AREX-2 takes a different route. The authors built their training data around a long "attempt, analyze, fix" cycle. In it, the model isn't limited to a single pass — it returns to its own solution, identifies the weak point, and reworks the algorithm.

The domain wasn't chosen at random. In ML engineering and algorithmic programming, results are measurable: code either passes tests or it doesn't, a metric either rises or falls. That makes the feedback reliable, and it allows long improvement trajectories to be collected and verified.

Why the 27B vs. 1.6T comparison grabs attention

The study's headline rests on a contrast in scale. If the claimed result holds up under independent benchmarking, it would reinforce a trend that's been building for a while: performance on narrow, well-verifiable tasks increasingly depends on how data is structured and how training is run, rather than on parameter count.

For engineers, this has practical implications. A model at the 27B scale 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 frame the work around recursive self-improvement (RSI) — a system's ability to iteratively improve its own algorithms. The idea is that a model capable of sustained, thoughtful refinement of code for ML tasks gradually moves closer to improving ML processes themselves. In their framing, this opens a door to autonomous evolution of intelligence. For now, this is a first step in a narrow domain, not a fully self-improving system.

Key release parameters:

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

What agent-building teams should verify

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

Independent benchmarking is needed before adoption. First, it's important to determine on which benchmarks and under what conditions AREX-2 beats the larger model — domain narrowness could heavily skew the conclusions. Next, the gains need to hold up on out-of-distribution tasks and over long horizons where errors accumulate. If those checks pass, compact models trained on reflection could become a real alternative to expensive giants for engineering agents.