{"id":"halcyon-radical-numerics","slug":"halcyon-radical-numerics","name":"Halcyon","orgId":"radical-numerics","family":"halcyon-radical-numerics","releaseDate":null,"availability":"announced","openWeights":true,"license":null,"paramsTotal":1700000000,"paramsSource":"slug","contextWindow":null,"knowledgeCutoff":null,"modelCardUrl":"https://x.com/RadicalNumerics/status/2110766682977521742","hfRepo":null,"modalities":["image"],"aliases":["Halcyon","radical numerics halcyon","halcyon-4b","halcyon-1.7b","radicalnumerics/halcyon-4b"],"aliasNote":null,"origin":"tracked","tracked":{"trackedSince":"2026-10-21","announced":"2026-10-15","blurb":"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.","api":"none","openWeights":"yes","params":{"text":"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)","source":"Hugging Face model cards"},"facts":[{"label":"Lab","value":"Radical Numerics, with Stanford University (paper authors include Li Fei-Fei and Michael Poli)"},{"label":"Architecture","value":"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"},{"label":"Training data","value":"One epoch of GPIC, 100M image-text pairs at 256x256 (97,632 steps at batch 1,024); no text-only pretraining"},{"label":"Output","value":"256x256 images as 1,024 tokens (32x32 grid); the I2T variant also captions images"},{"label":"Release","value":"Ten checkpoints: five sizes, each with a main (T2I + 10% image-to-text) and a baseline branch, plus the tokenizers"},{"label":"Access","value":"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"}],"claims":["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":[{"kind":"lab","title":"Radical Numerics announcement on X","url":"https://x.com/RadicalNumerics/status/2110766682977521742","date":"2026-10-15"},{"kind":"lab","title":"Text-to-image generative modeling without language pretraining (Radical Numerics blog)","url":"https://www.radicalnumerics.ai/blog/halcyon","date":"2026-10-15"},{"kind":"repo","title":"Halcyon models on Hugging Face","url":"https://huggingface.co/radicalnumerics","date":"2026-10-15"},{"kind":"paper","title":"Text-to-Image Generative Modeling Without Language Pretraining (arXiv 2610.26863)","url":"https://arxiv.org/abs/2610.26863","date":"2026-10-19"}],"firstBenchmark":null},"organization":{"id":"radical-numerics","name":"Radical Numerics","slug":"radical-numerics","country":null,"url":"https://www.radicalnumerics.ai","color":"#11110F","mark":"R"},"price":null,"answer":{"word":"Pending benchmarks.","scope":"Announced Oct 15, 2026; no independent results yet.","lead":"As of Oct 22, 2026, Halcyon is pending benchmarks on isitsota: announced Oct 15, 2026, with no independent results on the boards we track yet.","confidence":"low"},"status":"PENDING_BENCHMARKS","verdicts":[]}