Axion 1.0: Open Agent Foundation Models, Simplified

Axion 1.0 is a fully open agent foundation model trained with two SFT stages plus RL, without dedicated mid-training, DPO, or OPD, and is competitive across a broad agentic benchmark suite.

Modern agent models often rely on continued or mid-training, multiple supervised stages, specialized capability-specific data mixtures, preference optimization, RL, and additional post-training, making it difficult to identify which ingredients are necessary and raising reproduction cost. Many open models release weights while leaving data construction, filtering, verification, prompts, and training procedures unavailable. The paper asks how much of this complexity is actually necessary to build a strong and broadly capable agent foundation model.

Axion 1.0 uses three stages: two SFT stages followed by RL, without dedicated mid-training, DPO, or OPD. Stage I constructs self-contained reasoning tasks from public seeds; a generator composes new questions, the base model is evaluated with Python but no search, and problems it solves are strengthened by composition, parameterization, range shift, or case analysis, with reference answers accepted only when three independent GPT-5.5 solutions agree. Stage II builds tool-dependent tasks over structured public data using deterministic filter, select, compute, and verbalize steps; construction-time databases are withheld during solving and closed-book-solvable questions are removed. SFT serializes full interactions but supervises only assistant tokens, and Stage II is initialized from the Stage I checkpoint. RL starts from Stage II and uses GRPO with final-answer and intermediate-progress rewards, selecting tasks where four sampled trajectories include a correct answer and have nonzero reward variance.

• On full HLE with restricted website access, Axion 1.0-35B and Axion 1.0-122B achieve 52.2% and 59.0%, respectively; without these restrictions, the corresponding scores are 59.1% and 64.8%. §4.2
• Compared with similarly sized Agents-A1, Axion 1.0-35B improves by 4.6% on HLE (52.2% vs. 47.6%), 5.8% on SEAL-0 (62.2% vs. 56.4%), and 7.0% on FrontierScience Olympiad (86.0% vs. 79.0%). §4.2
• On FrontierScience Research, the 35B and 122B models score 56.7% and 66.7%, compared with 40.0% for Agents-A1 and 26.7% for GPT-5.5; the 122B model exceeds GPT-5.6 Sol’s 63.3% by 3.4%. On BMB, Axion 1.0-122B reaches 41.7%, compared with 47.1% for GPT-5.6 Sol and 46.1% for Claude Fable 5. §4.2
• Increasing model size from 35B to 122B improves five of six benchmark scores, with largest gains on FrontierScience Research (+10.0%), HLE (+6.8%), and BMB (+6.4%), and a smaller gain on HiPhO (+0.5%); SEAL-0 is the exception, at 62.2% for the 35B model and 58.6% for the 122B model. §4.2
• For the 35B model, SFT Stage I raises HLE-130 from 47.7% to 59.2%, HiPhO from 38.6% to 45.9%, and FrontierScience Olympiad from 71.0% to 83.0%, while BMB decreases from 23.5% to 17.6%; SFT Stage II then raises BMB from 17.6% to 35.3%, FrontierScience Research from 48.3% to 58.3%, SEAL-0 from 50.5% to 53.2%, and HLE-130 from 59.2% to 66.2%. §4.3
• For the 35B model, RL increases HLE-130 from 66.2% to 70.0% and SEAL-0 from 53.2% to 62.2%; for the 122B model, they increase from 68.5% to 74.6% and from 53.2% to 58.6%, respectively, and BMB improves from 35.3% to 41.7%. §4.3

• Cross-source baseline comparisons may involve different tool access or execution settings and are described as indicative rather than strictly controlled (§4.1).
• Training-stage comparisons follow sequential checkpoints and are not component-removal ablations (§4.3).
• The 35B and 122B results are not a controlled size-scaling experiment because these models additionally undergo RL and use different base architectures (§4.4).
• Benchmark differences under rollout routing replay cannot be causally attributed to R3 alone because the log-probability diagnostics also reflect asynchronous policy staleness and off-policy context-condensation effects (§4.4).

• Axion 1.0 is released with the complete recipe, including data construction, generation, filtering, verification, prompts, and training code, not only model weights. §Abstract
• The training recipe has only three stages: two lightweight SFT stages followed by RL, with no dedicated mid-training, DPO, or OPD. §3.1
• Stage I uses model-aware data construction from public seeds; problems the base model solves are strengthened with composition, parameterization, range shift, or case analysis, and reference answers are accepted only when three GPT-5.5 solutions agree. §2.1
• Stage II constructs tool-dependent questions through filter, select, compute, and verbalize steps over structured data; construction-time databases are withheld and closed-book-solvable questions are removed. §2.2
• SFT supervises only assistant-generated tokens—reasoning, tool calls, and final answers—while system and user messages and tool observations provide context without contributing to the loss. §3.2
• RL gives full reward for a correct final answer; otherwise partial credit is max(0, 0.3P - 0.1(1-A)), with progress P aggregated in dependency order so later steps count only after prerequisites are verified. §3.3
• All SFT runs use a 128K context window on 32 GPUs; Qwen3.5-122B-A10B takes 7 h 30 min and 240 GPU-hours total, while Qwen3.6-35B-A3B takes 2 h 48 min and approximately 90 GPU-hours. §5.1
• Async RL separates rollout and training GPU pools; TITO preserves rollout token IDs, and R3 records and replays MoE expert routing, reducing mean gradient norm from 1.998 to 1.922 and mean PPO KL from 0.0413 to 0.0396. §5.2