Axion 1.0
A 35B-parameter fully open-source agent foundation model released by ByteDance. It employs a minimalist two-stage training recipe of SFT plus reinforcement learning, skipping mid-training, DPO, and OPD, and emphasizes verifiable rewards and tool-calling capabilities.
Timeline (last 90 days)
The ByteDance team released Axion 1.0, a fully open 35B-parameter agent foundation model, and simultaneously released the paper, code, datasets and project homepage.
Axion 1.0's training recipe has only three stages: two lightweight SFT stages and one reinforcement learning stage.
Axion 1.0's training pipeline does not include dedicated mid-training, direct preference optimization (DPO) or on-policy distillation (OPD).
The first SFT stage adopts a "model-aware data construction" method: first synthesizing challenging problems, then adversarially strengthening these problems based on the base model's own successful solving strategies.
The second SFT stage built a scalable data pipeline for constructing diverse and automatically verifiable tasks, based on real data sources and deterministic computation.
The reinforcement learning stage aims to further strengthen long-horizon reasoning ability as well as the model's ability to interact with external tools and environments.
Officials say Axion 1.0 achieved strong performance on a broad set of agent benchmarks and is competitive with leading commercial models on several of them.
The official materials do not provide specific benchmark scores or comparison figures.
Axion 1.0's openness extends beyond the model weights, with the complete recipe released as well, including data construction, generation, filtering, validation, prompts, and training code.
All Facts (9)
Axion 1.0's openness extends beyond the model weights, with the complete recipe released as well, including data construction, generation, filtering, validation, prompts, and training code.
50%UnverifiedAxion 1.0's training recipe has only three stages: two lightweight SFT stages and one reinforcement learning stage.
50%UnverifiedAxion 1.0's training pipeline does not include dedicated mid-training, direct preference optimization (DPO) or on-policy distillation (OPD).
50%UnverifiedThe first SFT stage adopts a "model-aware data construction" method: first synthesizing challenging problems, then adversarially strengthening these problems based on the base model's own successful solving strategies.
50%UnverifiedThe second SFT stage built a scalable data pipeline for constructing diverse and automatically verifiable tasks, based on real data sources and deterministic computation.
50%UnverifiedThe reinforcement learning stage aims to further strengthen long-horizon reasoning ability as well as the model's ability to interact with external tools and environments.
50%UnverifiedOfficials say Axion 1.0 achieved strong performance on a broad set of agent benchmarks and is competitive with leading commercial models on several of them.
50%UnverifiedThe official materials do not provide specific benchmark scores or comparison figures.
50%UnverifiedThe ByteDance team released Axion 1.0, a fully open 35B-parameter agent foundation model, and simultaneously released the paper, code, datasets and project homepage.
50%