An AI agent able to search the web, run code and complete multi-step tasks may come out of a much shorter training path than assumed. That is the argument of researchers led by Bytedance, TikTok's parent company, together with several US universities. Their model, Axion 1.0, is said to hold its own against top commercial systems from OpenAI and Anthropic on benchmarks, despite being trained in just three phases.
Fewer phases, better data
The paper argues that agent training pipelines have become too layered. Developers stack one phase after another (mid-training, supervised fine-tuning, preference optimisation, online distillation, reinforcement learning) and, according to the authors, it has become almost impossible to tell which step actually helps. Reproducing the work is also getting expensive.

Image: OpenRouter — https://openrouter.ai/compare/aion-labs/aion-1.0
Axion 1.0's recipe is simpler: two supervised fine-tuning stages followed by reinforcement learning, with no mid-training, DPO or online distillation. The researchers credit the results to careful selection of training data rather than to extra training steps. The team has also released the code used to build, generate, filter and verify its datasets, along with prompts and training code.
The claimed numbers
In the paper, the 122-billion-parameter version scores 59.0% on a variant of Humanity's Last Exam with restricted web access, and 66.7% on FrontierScience Research. The 35-billion-parameter version reaches 62.2% on SEAL-0, a web information-seeking test. These figures come from the authors and have not yet been checked by independent evaluations.
The results should be read carefully. On the Humanity's Last Exam reasoning test, Axion stays behind several recent closed models, while on other benchmarks it is competitive or ahead. The picture is of a model close to the best, not uniformly better.
The model, dataset and code are freely available. For a startup building agents, the takeaway is that investing in data quality may pay off more than lengthening the pipeline, with lower compute and engineering costs. Because everything is public, small teams can test the method instead of relying only on the tables.
The work fits a broader trend: open models from Chinese and US labs are closing in on closed systems, and competition is shifting toward the balance between cost and performance. The next test will be whether the results hold up in independent evaluations.

No comments yet. Be the first!