Halcyon
Stanford and Radical Numerics' text-to-image models trained from scratch with no language pretraining, using up to 275x less compute.
Overview
Halcyon is a family of early-fusion autoregressive text-to-image models from Stanford researchers Keshigeyan Chandrasegaran, Fei-Fei Li and Jiajun Wu with Radical Numerics, published October 20, 2026. The two sizes, Halcyon-0.52B and Halcyon-1.6B, are trained from scratch on GPIC, a 100M image-text dataset, with no language-model pretraining. On the paper's gFD metric, Halcyon-0.52B scores 48.0 against 63.3 for LlamaGen-XL/8 at 18.2x less pretraining compute, and Halcyon-1.6B scores 35.3 against 41.0 for JiT-T2I-XXL/8++ at 275x less compute. Controlled experiments in the paper indicate text-only pretraining does not improve text-to-image generation. License and weight availability are not stated.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Two sizes: 0.52B and 1.6B parameters
- Early-fusion autoregressive design, no language pretraining
- Trained on GPIC, a 100M image-text dataset
- gFD 48.0 (0.52B) and 35.3 (1.6B) versus 63.3 and 41.0 for compared baselines
- 18.2x and 275x less pretraining compute than those baselines
Use cases
- Research on compute-efficient image generation
- Baselines for studying whether language pretraining helps image synthesis
Pricing
unknown; research release, license and weights not stated.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 20 Oct 2026
- Record updated
- 28 Oct 2026
- Record last checked
- 28 Oct 2026