The release is part of a research program at Radical Numerics focused on the science of multimodal pretraining and scaling. The study specifically tests whether text-to-image generation requires language pretraining to achieve competitive performance.
Tuesday, October 20, 2026
Story
Radical Numerics' Halcyon Models Cut Pretraining Compute 275x
AI5 days ago
A family of early-fusion base models ranging from 0.1B to 4B parameters can outperform late-fusion text-to-image baselines without using frozen language models or text-only pretraining. Developed by Radical Numerics as Halcyon, these models were trained from scratch on 100M image-test pairs from GPIC. The 1.6B Halcyon model beats the JiT-T2I-XXL/8++ baseline while using 275x less pretraining compute, and the 0.52B version outperforms LlamaGen-XL/8 with 18.2x less compute, including the cost of the language models used by the baselines. Radical Numerics has made the models and their training recipes open.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.