Axion 1.0 open agent foundation model released

Axion 1.0 is a 35B fully open agent foundation model trained with only two supervised fine-tuning stages plus reinforcement learning, skipping mid-training, DPO, and OPD. The team released the full recipe including data construction, verification, prompts, and training code, claiming competitiveness with leading commercial models on several agentic benchmarks.

• Protegrity releases open-weights 278M PII detection model and benchmark2026-11-04
• Palette fine-tunes 20B model to build decks in brand design language2026-11-03
• DeepSeek's $15.9B round lifts October global venture funding to $59B2026-11-03
• Transformers adds small model fixtures to catch regressions2026-11-03
• Hugging Face releases FineDNA dataset to teach LLMs DNA2026-11-01