Is Halcyon state of the art?Pending benchmarks.Announced Oct 15, 2026; no independent results yet.
Radical Numerics' family of early-fusion autoregressive text-to-image base models, trained entirely from scratch on 100M image-text pairs (GPIC) with no language pretraining: one decoder-only Transformer over a joint text-image vocabulary, with its own text and image tokenizers. Five sizes from 0.1B to 4B, each in two training variants, on Hugging Face (gated, no license stated). Research release with Stanford; 256x256 images.
The details so far
- Lab
- Radical Numerics
- Announced
- Oct 15, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- Five sizes: 96M, 248M, 523M, 1.57B and 3.84B (paper names 0.1B, 0.25B, 0.52B, 1.6B, 4B; Hugging Face repos Halcyon-0.1B, -0.25B, -0.6B, -1.7B, -4B)
- Lab
- Radical Numerics, with Stanford University (paper authors include Li Fei-Fei and Michael Poli)
- Architecture
- Early-fusion decoder-only Transformer (Qwen3 architecture, random init), next-token prediction over 80,640 tokens: 16,384 text BPE + 64,000 image codes + specials
- Training data
- One epoch of GPIC, 100M image-text pairs at 256x256 (97,632 steps at batch 1,024); no text-only pretraining
- Output
- 256x256 images as 1,024 tokens (32x32 grid); the I2T variant also captions images
- Release
- Ten checkpoints: five sizes, each with a main (T2I + 10% image-to-text) and a baseline branch, plus the tokenizers
- Access
- Hugging Face, gated (access request); no license stated on the model cards. No Halcyon training-code repository found; recipes are documented in the paper and cards
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: Halcyon-1.6B beats the late-fusion JiT-T2I-XXL/8++ (35.3 vs 41.0 gFD on GPIC, lower is better) with 275x less total pretraining compute; Halcyon-0.52B beats LlamaGen-XL/8 (48.0 vs 63.3) with 18.2x less.
Lab claim: in its controlled experiments, adding text-only pretraining does not improve text-to-image generation.
Lab claim: 4B reaches 28.4 gFD and 56.39 GenEval; 1.6B scores 53.48 GenEval and 77.90 DPG-Bench.
Sources
LABRadical Numerics announcement on Xx.com/RadicalNumerics/status/2110766682977521742 · Oct 15, 2026LABText-to-image generative modeling without language pretraining (Radical Numerics blog)radicalnumerics.ai/blog/halcyon · Oct 15, 2026REPOHalcyon models on Hugging Facehuggingface.co/radicalnumerics · Oct 15, 2026PAPERText-to-Image Generative Modeling Without Language Pretraining (arXiv 2610.26863)arxiv.org/abs/2610.26863 · Oct 19, 2026Tracking continues until an independent board lists it.