ByteDance and partners open-source Axion 1.0, a 35B agent base model
11/04/2026 — 11/05, 01:19·2 sources (all)·2 reports (all)
Key takeaways
- ByteDance and Duke University researchers have open-sourced Axion 1.0, a fully open 35B agent base model built with a three-stage training r…
Background & analysis
On November 4, 2026 at 17:02, a post appeared on Reddit's LocalLLaMA subreddit introducing Axion 1.0, a fully open agent foundation model with 35B parameters. According to the post, Axion 1.0 was trained in only three stages: two lightweight supervised fine-tuning (SFT) stages followed by reinforcement learning (RL). The pipeline contains no dedicated mid-training phase, no direct preference optimization (DPO), and no on-policy distillation (OPD). The post explains that the first SFT stage develops strong self-contained reasoning through model-aware data construction, in which reasoning data is synthesized. As of posting time, no further details about the model were provided.
AI-generated from 1 reports · updated 1 day ago
Latest turnA team from ByteDance, Duke University and other institutions released Axion 1.0, a fully open Agent foundation model trained on 18,363 trajectories with a three-stage recipe. It reports competitive results on hard scientific reasoning and tool-use benchmarks, and the 122B variant took roughly 1,000 GPU-hours; everything is released under Apache 2.0.
- Heat index
- 83
- All sources
- 2
- All reports
- 2
- First seen
- 1 day ago
Reports on this story headlines open the original
A team from ByteDance, Duke University and other institutions released Axion 1.0, a fully open Agent foundation model trained on 18,363 trajectories with a three-stage recipe. It reports competitive results on hard scientific reasoning and tool-use benchmarks, and the 122B variant took roughly 1,000 GPU-hours; everything is released under Apache 2.0.
DEV Community · AIAI score 64
Axion 1.0 is a fully open 35B agent foundation model trained in just three stages: two lightweight supervised fine-tuning rounds followed by reinforcement learning, with no dedicated mid-training, DPO, or on-policy distillation. The first SFT stage builds self-contained reasoning through model-aware data construction.
Reddit · LocalLLaMAAI score 67
Other stories people are talking about
- Heat index 840OpenAI ships Images 2.6 with GIF and SVG output15 sources
- Heat index 675OpenAI says its cyber model pipeline confirmed 1,724 Linux kernel flaws and generated 948 privilege-escalation exploits11 sources
- Heat index 642RisingOpenAI GPT-5.6 Brings Major Agentic and Computer Use Upgrades13 sources
- Heat index 633RisingCat Wu, product lead for Claude Code, joins OpenAI9 sources
- Heat index 630NewSurgeOpenAI releases GPT-6 Cyber cybersecurity model7 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion