---
format: "aidr-story-markdown/v1"
id: "32592320620553c190ad69f19bdc3ecf0a613bd6c14481b7214f0b45e1d6baf5"
canonical_url: "https://aidr.today/32592320?lang=en"
title: "Radical Numerics' Halcyon Models Cut Pretraining Compute 275x"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-15T20:04:40.000Z"
category: "Open Source"
topics: ["open-source"]
source_urls: ["https://marketbrief.now/ai/radical-numerics-halcyon-models-cut-pretraining-compute-275x-821c8f32","https://huggingnews.com/ai/radical-numerics-halcyon-models-cut-pretraining-compute-275x-821c8f32"]
summary: "Radical Numerics' Halcyon 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. 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. Radical Numerics has made the models and their training recipes open."
---

# Radical Numerics' Halcyon Models Cut Pretraining Compute 275x

> [Open the canonical story](<https://aidr.today/32592320?lang=en>)

**Published:** 2026-10-15T20:04:40.000Z
**Category:** Open Source
**Topics:** open\-source

## Summary

Radical Numerics' Halcyon 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\. 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\. Radical Numerics has made the models and their training recipes open\.

## Sources

- [Story source](<https://marketbrief.now/ai/radical-numerics-halcyon-models-cut-pretraining-compute-275x-821c8f32>)
- [Story source](<https://huggingnews.com/ai/radical-numerics-halcyon-models-cut-pretraining-compute-275x-821c8f32>)