Axion 1.0

Axion 1.0 is an open agent foundation model family from ByteDance, described in the paper Axion 1.0: Open Agent Foundation Models, Simplified (HF Papers 2611.03233, posted 2026-11-02), and widely discussed on 2026-11-05.

It comes in two main configurations:

• Axion 1.0-35B (35B total / 3B active, initialized from Qwen3.6-35B-A3B)
• Axion 1.0-122B (122B total / 10B active, initialized from Qwen3.5-122B-A10B)

The paper also reports SFT-only 4B and 9B variants. The headline is simplicity: roughly 18K SFT examples, two SFT stages plus one RL stage, and no dedicated mid-training, DPO or on-policy distillation, yet competitive results on agentic science and search benchmarks such as HLE, SEAL-0, FrontierScience, BioMysteryBench and HiPhO. Code and recipe live at GitHub bytedance/axion1_0; a community reproduction of the 35B checkpoint is on Hugging Face at Axion1-0/Axion-1.0-35B (Apache 2.0).

What problem it solves

Building a research-style agent model that can search, browse, run Python and query scholarly sources usually requires heavy pipelines: mid-training, DPO, on-policy distillation. Axion 1.0 argues that a small amount of verifiable, high-quality data plus a simple three-stage recipe is enough to be competitive.

Key numbers (official Table 1)

Model	HLE	SEAL-0	FS Olympiad	FS Research
Axion 1.0-35B	52.2	62.2	86.0	56.7
Axion 1.0-122B	59.0	58.6	87.0	66.7

• vs. the similarly sized Agents-A1, the 35B gains +4.6 on HLE and +5.8 on SEAL-0.
• The 122B scores 66.7 on FrontierScience Research vs. 63.3 for GPT-5.6 Sol (team-run evaluation).

Recipe

1. Two SFT stages (full-parameter, sample packing).
2. Asynchronous GRPO RL with 8 rollouts per question and session-level training across context-condensation events.
3. Reasoning problems hardened using base-model feedback; tool tasks built from public data with programmatic reference answers.

Caveats

• The official repo mainly ships the paper and recipe; the 35B weights on HF are a community reproduction, not an official checkpoint.
• Cross-source baselines use different tools/settings, so comparisons are indicative.
• It targets scientific reasoning and information seeking, not general chat or coding.

Who it's for

• Research teams reproducing or training research/search agent models;
• Developers who want a deep-research agent backbone with ~3B active parameters that can run on their own hardware.

Getting started

1. Read the project page axion1-0.github.io and the paper PDF.
2. Browse GitHub bytedance/axion1_0 for data filtering, prompts, SFT/RL and evaluation scripts.
3. Try the community reproduction: AutoModelForCausalLM.from_pretrained("Axion1-0/Axion-1.0-35B") or serve it with vLLM inside a ReAct harness with search, page-visit and Python tools.
4. When benchmarking, restrict website access as the authors do (e.g. block Hugging Face) to avoid answer leakage.