Halcyon: Text-to-Image Base Models That Cut Pre-training Compute 275x

Radical Numerics, with Stanford AI Lab researchers, released Halcyon, open-source early-fusion text-to-image base models trained from scratch without frozen language models, saving up to 275x pre-training compute; a 0.52B variant beats LlamaGen-XL/8 with 18x less compute.

2026-10-16 ~ 2026-10-16 · 3 related posts

• Halcyon early-fusion image models match late-fusion with 275x less pretraining compute, open-sourced — davidwromero · 2026-10-16
• Halcyon: no text pretraining, beats LlamaGen-XL/8 with 18× less compute — JosephJacks_ · 2026-10-16
• Halcyon: from-scratch early-fusion T2I models match late fusion with 275x less pretraining compute — StanfordAILab · 2026-10-16