Axion-1.0-35B

Axion-1.0-35B

A reproduction checkpoint of Axion 1.0-35B, a 35B mixture of experts model for scientific reasoning, information seeking and tool using problem solving, served with vLLM. Suits researchers and agent developers.

Made by Axion1-0 on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository offers a rebuilt copy of Axion 1.0-35B, trained with the method from its paper. It is a mixture of experts model with about 35B parameters in total and 3B active, stored in BF16. It is meant for scientific reasoning, finding information and solving problems with tools.

Who it suits: Researchers and engineers who build agents that reason through science questions or look things up with tools, and anyone who wants to check the paper's results with an open checkpoint.

Industry | Who uses it | How they use it | Example
Research | Science and AI researchers | Reproduce and test the paper's agent recipe | Run the checkpoint on benchmark questions
Software | Agent developers | Use it as the reasoning core of a tool using agent | Serve it with a tool call parser
Education | Advanced students | Study a three stage training recipe | Compare the paper scores with the reproduction

Skip it if you have no multi GPU server, because the card's serving example uses eight GPUs.

Where to use it:

• A GPU server running vLLM, as in the card's example
• An agent runner that carries out tool calls, which the card says tool using evaluation needs
• A Python setup with the Hugging Face tools, to fetch the files

Use case | Who | What they get back
Science reasoning | Researchers | Worked answers to hard science questions
Information seeking with tools | Agent builders | Answers that rely on tool calls
Chat through a served endpoint | Developers | Replies from the model over an OpenAI compatible interface

How to set it up:

• Python and the Hugging Face tools
• vLLM
• Several GPUs, since the example uses eight and you can adjust the tensor parallel size and context length to your memory
• About 70.2 GB of disk for the BF16 weights, as an estimate

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Serve the checkpoint with vLLM, using the card's command.

pip install -U huggingface_hub
hf download Axion1-0/Axion-1.0-35B

vllm serve Axion1-0/Axion-1.0-35B \
  --served-model-name axion-35b \
  --tensor-parallel-size 8 --dtype bfloat16 \
  --max-model-len 131072 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

How to use it: After the server starts, send chat requests to its OpenAI compatible endpoint, naming the model axion-35b. The card also sets a host and port, which you can choose yourself.

Tool | What it does
vllm serve | Starts the model server
--tensor-parallel-size | Spreads the model across GPUs, 8 in the example
--max-model-len | Sets the context length, 131072 in the example
--tool-call-parser | Reads tool calls out of the model's output

Example: Serve the model as shown and send a chat request with the user message "What is 2 + 3?" to axion-35b. The reply comes back as a normal chat completion. Tool using agent runs need a runner that executes the calls.

For newcomers and experts: If you are new to this, know that this is a large model that needs several GPUs. Start with the plain sum question above before you try agent work.

If you know the field, the recipe has three stages on about 18K supervised examples. First, supervised fine tuning on hard self contained science problems. Second, supervised tuning on verifiable tasks that need outside information and tools. Third, reinforcement learning on agent runs.

A question goes in
The model reasons step by step
A runner carries out any tool calls
The answer comes back

Title: Reproduction checkpoint scores
HLE: 52.8 percent
BMB: 35.3 percent
SEAL-0: 60.4 percent
HiPhO: 45.9 percent
FS Olympiad: 85.0 percent
FS Research: 55.0 percent

Good to know:

• The card compares the reproduction with the paper, and the two are close but not identical.
• HLE is Humanity's Last Exam, BMB is a BioMysteryBench subset and FS is FrontierScience.
• License: apache-2.0.
• Cost to run: none named, only your own GPU hardware.
• Data it reaches: the prompts you send and any tool output. The source names no account or outside service.
• Not for: running tools by itself. The card says agent evaluation needs a separate runner.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 35.1B
• Weights: about 70.2 GB in BF16; about 17.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,535
• Likes: 76
• License: apache-2.0
• Quantized versions: 8
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, reproduction, agent, tool-use, scientific-reasoning, text-generation, conversational

Links

• Model: https://huggingface.co/Axion1-0/Axion-1.0-35B
• Files: https://huggingface.co/Axion1-0/Axion-1.0-35B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Axion1-0/Axion-1.0-35B

Then follow the usage in its model card. Nothing in this kit runs by itself.