The community has reproduced and released a scientific reasoning model developed by ByteDance researchers and others. Developers can download and test a model that uses search and code execution. The accompanying materials let them examine its data composition and training process.
1. A Qwen-based community reproduction
Axion 1.0-35B on Hugging Face provides community-reproduced weights. The researchers' repository identifies Qwen3.6-35B-A3B as the paper model's base and links to the reproduction. Further training refines the base model's scientific reasoning and tool use.
The model card lists about 35 billion parameters, the internal values trained to produce answers, with about 3 billion active during processing. Its mixture-of-experts architecture selects some experts, using only some computation paths. The weights are available in BF16 format and marked with the Apache 2.0 license.
2. From learning solutions to search and interaction
The paper uses two stages of supervised fine-tuning, further training on example solutions, followed by reinforcement learning with rewards for tool use. First, the model learns to solve science and math problems using information within each problem. Next, it learns to search for missing information and perform calculations. Finally, it learns to complete answers using search, page browsing, academic search, and Python execution.
Data construction makes problems the base model already solves harder to train reasoning. Tool-use problems start by selecting targets from public data, calculating answers, and turning them into questions. Since the problem creators obtain answers first, they can verify solutions produced through search and calculation.
3. Tracking training-stage changes on the same evaluation
The paper reports results at each training stage for the same base model. Below are results on 130 text-only questions from Humanity's Last Exam, a challenging cross-disciplinary evaluation. The researchers used the highest single-run score from three runs.
Based on Qwen3.6-35B-A3B; 130 text-only HLE-130 questions; environment supporting search, page browsing, academic search, and Python, with evaluation-specific restrictions; highest single-run score from 3 runs
| Training stage | Score |
|---|---|
| Base model | 47.7 |
| After supervised fine-tuning stage 1 | 59.2 |
| After supervised fine-tuning stage 2 | 66.2 |
| After reinforcement learning | 70.0 |
Scores rose sharply across the two supervised fine-tuning stages, with reinforcement learning adding another gain. On FrontierScience, which tests research-level science problems, scores fell from 58.3% to 56.7% after reinforcement learning. This shows that training effects vary by task.
On the full Humanity's Last Exam, the researchers reported 52.2% with access to some websites restricted and 59.1% without those restrictions. The reproduction's model card lists 52.8% on the full exam and 60.4% on SEAL-0, an information-seeking evaluation.
The reproduction's card gives no separate measurement conditions, limiting direct comparisons with the paper's results. The evaluation guide also says GPU serving and the paper's benchmark results were not revalidated for this release.
4. Prepare model serving and tool connections separately
The model card provides vLLM serving instructions, for running the model to receive requests and return responses, and OpenAI-compatible request examples.
The command below is the card's original launch example, using 8 GPUs, BF16, and a maximum length of 131,072.
vllm serve Axion1-0/Axion-1.0-35B \
--served-model-name axion-35b \
--host 127.0.0.1 --port 8000 \
--tensor-parallel-size 8 --dtype bfloat16 \
--max-model-len 131072 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice --tool-call-parser qwen3_coderThe card advises adjusting parallelism, the number of devices sharing computation, and context capacity, the input range the model can consult at once, to fit GPU memory. Once the model is running, developers can send requests through an OpenAI-compatible API.
Search-based evaluation needs an external executor that connects model tool calls to actual operations. It connects search services, page browsing, and an isolated code execution environment, returning tool results in the next model request. The repository's evaluation code scores solutions collected by this external executor.
5. Open materials support experiments, but hardware costs remain
Released materials cover data generation, filtering, validation, prompts, training code, and evaluation preparation. Developers can trace the problems the model learned and how answers were checked. Beyond testing finished weights, they can more easily modify training procedures and reuse evaluation workflows.
Users still bear the cost of retraining hardware and operating tool services. The paper reports using a 128K context and 32 GPUs for the 35B model's two supervised fine-tuning stages. This took about 90 GPU-hours in total and up to 101GiB of memory per GPU. The open code provides a starting point for experiments, but running them requires substantial computing resources.
Axion 1.0-35B's value lies less in leaderboard rankings than in the ability to modify its open training methods and test them with reproduced weights.