Systems, scaling and architecture for general biological intelligence
San Francisco & Tokyo
Joined May 2025
- Tweets51
- Following4
- Followers4.3K
- Likes48
Launching Omnii-1, our most powerful genome language model yet.
Radical Numerics retweeted
I'm thrilled to welcome Jill Hruby @jill_hruby and Paul Friedrichs as biodefense advisors to @RadicalNumerics!
Jill formerly led the National Nuclear Security Administration, and Paul served as the inaugural White House Director of Pandemic Preparedness and Response.
As AI opens new frontiers in biology, strengthening our ability to defend against biological threats is more important than ever.
We're incredibly fortunate to learn from their experience and leadership as we advance this mission.
Welcome aboard, Jill and Paul!
Read the announcement:
radicalnumerics.ai/blog/biod…
1/8 Does text-to-image generation need language pretraining?
Today, we share Halcyon, a family of early-fusion, base models trained from scratch. Halcyon outperforms strong late-fusion baselines with up to 275× less pretraining compute. Models and recipes are open.
Halcyon is part of a research program at Radical Numerics on the science of multimodal pretraining and scaling.
8/8 Classifier-free guidance needs an unconditional image branch, usually trained by dropping captions. Halcyon instead places the caption after the image. The same model can then caption images at no extra pretraining cost, while keeping competitive generation quality.
read more: radicalnumerics.ai/blog/halc…
paper: radicalnumerics.ai/assets/ha…
website: cs.stanford.edu/~keshik/halc…
We're hiring: radicalnumerics.ai/join-us
Work done by @keshigeyan at @RadicalNumerics, @StanfordAILab and @StanfordSVL, with @__yuezhao__, @athmsx, @eadeli, @jcniebles, @jiajunwu_cs, @drfeifei and @MichaelPoli6.
6/8 We test whether language pretraining helps in three settings. A larger language encoder gives a marginal gFD gain at 13.5× the language pretraining compute. Text-only pretraining does not improve gFD. Continual pretraining from Qwen3, does not improve over training from scratch.
4/8 Text-to-image models are usually compared on shared benchmarks after pretraining on different datasets, which conflates data and method. We retrain LlamaGen and JiT-T2I on the same 100M GPIC pairs at 256×256, include stronger variants of each baseline, and evaluate every model under one protocol.
2/8 Halcyon includes no text-only pretraining and no frozen language models.
At 0.1-4B parameters, trained on 100M image-test pairs from GPIC, it outperforms strong late-fusion baselines. The 0.52B model outperforms LlamaGen-XL/8 with 18.2× less pretraining compute, and the 1.6B model outperforms JiT-T2I-XXL/8++ with 275× less. Compute totals include the pretraining of the language models used by the baselines.
1/5 What does it take to scale pretraining to 20+ data modalities at once? Improved data mixing methods are one important ingredient.
Today, we share how we scale data mixing to thousands of sources for DNA and natural language pretraining of Omnii, our multimodal genome model.
5/5 The same pipeline can be applied to multimodal datasets, from web text to DNA. DNA sources are usually taxonomic groups or the source dataset.
With Tessera, we can instead group the samples by what each sequence teaches the model.
Building general biological intelligence requires advanced data mixing methods at every stage of training, across modalities and scales.
Read more: radicalnumerics.ai/blog/tess…
We're hiring: radicalnumerics.ai/join-us
Welcoming Christine Parthemore to Radical Numerics as Head of Biodefense!
I'm excited to announce that Christine Parthemore @clparthemore has joined @RadicalNumerics as head of biodefense.
At Radical, we have a dual-mandate to drive the frontier of biological AI for both design *and* defense. Christine will lead the charge on making the best of our tech available to the biodefense community.
Christine was most recently head of the Council on Strategic Risk, and has been a pillar of the biodefense community in the US, Europe and Asia.
We're excited for her to drive our partnerships with allies around the world.
More on Christine's thoughts here:
radicalnumerics.ai/blog/chri…
Welcoming Jacob Rinaldi, Head of Scientific Strategy at Radical Numerics.
I'm excited to announce that Jacob Rinaldi has joined @RadicalNumerics as Head of Scientific Strategy!
Jacob was most recently cofounder of Noetik, leading their ML efforts.
Previously he was at Recursion as Head of Oncology, and Genentech, working on cancer vaccines.
At @RadicalNumerics, he'll work to accelerate the application of our models to some of biology’s most important problems, from human genetics and rare disease to cancer vaccines, whole-genome design, and the next generation of AI-enabled therapeutics.
Check out our announcement here:
radicalnumerics.ai/blog/jaco…
1/5 We pretrained Omnii, a multimodal language model on DNA. Then, we asked it to design a cancer vaccine
In new research, we show how Omnii learned to turn a tumor DNA sequence into a personalized mRNA vaccine design.
Read more here: radicalnumerics.ai/blog/omni…
4/5 Omnii predicts both for MHC class I and class II, ranks the best targets, then designs an optimized mRNA cassette.
On public benchmarks, Omnii outperformed specialized models on both antigen presentation and immunogenicity prediction, requiring only minimal post-training and no architectural modifications:
5/5 Moderna’s melanoma cancer vaccine just succeeded in Phase 3. Can we extend this to any cancer, any patient?
We are looking for researchers and partners who want to push this further.
Read the technical post:
radicalnumerics.ai/blog/omni…
