Researchers at ByteDance and three American universities have released Axion 1.0, an open agent model built with a simple recipe. Trained in only three stages, the largest version scores 59.0 on Humanity’s Last Exam, close to leading commercial systems. The team published the training recipe, data methods and code.
A Three-Stage Recipe, Nothing Extra
The recipe runs in three stages. First, supervised fine-tuning builds self-contained reasoning from 10,819 trajectories of hard science and maths problems, strengthened wherever the base model already succeeded. Second, another fine-tuning stage adds tool-dependent reasoning from 7,544 trajectories grounded in real data sources. Third, reinforcement learning sharpens long-horizon tool use.
What the recipe leaves out is the point: no dedicated mid-training, no direct preference optimisation, and no on-policy distillation. The authors argue that stacking training phases makes it nearly impossible to tell which step actually helps. Their answer is to put the difficulty into the data instead.
What the Benchmarks Show
Axion 1.0 comes in two sizes, 35 billion and 122 billion parameters, tested on six benchmarks for reasoning, search and science. The 122B model scores 59.0 on Humanity’s Last Exam, 87.0 on the FrontierScience Olympiad split and 66.7 on Research, while the 35B version scores 52.2, 86.0 and 56.7.
The 122B model beats GPT-5.6 Sol on FrontierScience Research, 66.7 to 63.3, and tops Gemini 3.1 Pro on Humanity’s Last Exam, 59.0 to 51.4. It trails Claude Fable 5 on that exam, 63.8 to 59.0, while the smaller 35B model leads the field on the SEAL-0 information-seeking test with 62.2.
Every figure carries the authors’ caveats. The scores are their evaluations, reported as the best of three runs rather than an average, and baseline comparisons are described as indicative rather than strictly controlled. The exam runs blocked websites, including Hugging Face; without that restriction the 122B score rises to 64.8.
Our Read
One fact keeps the story honest: Axion is built on Alibaba’s Qwen base models, not a ByteDance base model. The 35B version starts from Qwen3.6 and the 122B from Qwen3.5, so the achievement is post-training, not pretraining. That distinction shows how far a good recipe can carry somebody else’s foundation.
The timing adds a second layer. Axion skips on-policy distillation in the same week OpenAI published its defences against adversarial distillation. ByteDance is betting the opposite way: open recipes and simple training instead of guarding reasoning traces. The open release lets others test which approach holds up.
The release may matter more than any score. With the model, the team published its data construction, filtering and verification, plus prompts and training code. That means other teams can rerun the process instead of taking the benchmark table on trust, which is the reproducibility gap the paper complains about.
Why It Matters
If three stages suffice, smaller labs can build competitive agents without the pipelines only frontier labs can afford. The paper’s message is that data quality, not pipeline length, is the binding constraint on agentic capability. For ByteDance, Axion is a play in the open-model race, alongside Moonshot’s open Kimi weights.
Background
The paper was submitted to arXiv on 2 November and revised on 11 November; The Decoder covered it on 21 November. The work comes from ByteDance researchers with Duke University, the University of Maryland and the University at Buffalo, and both sizes use mixture-of-experts designs with a 128,000-token context window.
Source: the Axion 1.0 paper on arXiv and The Decoder’s report.


0 Comments