Radical Numerics

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. This was a pretty surprising emergent capability. Omnii was pretrained largely through next-token prediction on DNA, not for cancer vaccines. Post-trained, it could run the full pipeline: tumor DNA → neoantigens → presentation → immunogenicity → optimized mRNA vaccine A personalized cancer vaccine teaches the immune system which parts of a tumor to attack. The hard part is picking the right targets. Most tumor mutations are not useful vaccine targets. A candidate must pass two tests: will the tumor present it, and will the immune system respond? 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. Moderna's melanoma cancer vaccine just succeeded in Phase 3. Can we do this for any cancer, any patient? We're looking for researchers and partners who want to push this further. Read the technical post: https://lnkd.in/guVArrtt

Leaders of OpenAI, Anthropic, & GDM signed a letter urging Congress to mandate DNA synthesis screening. But chatbots & agents can't read DNA. So we built Omnii to defend against natural & AI-designed pathogens. 🖥️ Blog: https://lnkd.in/gx9stCYb 📃 Letter: screendna.org Today anyone can design DNA & get it shipped like an Amazon order from synthesis companies. This powers real science, but it's a biosecurity risk if bad actors design threats. AI raises the stakes: our prior model, Evo, generated the first AI-designed phage genome Our frontier genome language model, Omnii, reads and writes DNA/RNA and also understands proteins. It's function-aware: it screens whole genomes, not just genes/proteins, & flags threats with dangerous function even when unlike known pathogens. Ergo: generalized detection. Today's screening asks: "have we seen this before?" — aligning a sequence against known-threat databases. But sequence ≠ function. Distant proteins can fold alike, like dengue & chikungunya glycoproteins. Alignment misses this, and so we must screen beyond proteins. The hard cases are "paraphrases": same dangerous function, different sequence. Evolution designs them; a generative model designs them on demand. The kicker: a 442-residue protein can have ~10¹⁵⁶ viable variants. Enumeration is intractable. It's the blind spot of current tools. Omnii closes the gap: pathogenic functional signals live in its embedding geometry. Paraphrased pathogens cluster with their pathogenic templates, not harmless lookalikes, even when no alignment tool flags the link. A linear probe separates the classes at AUROC 0.991. Across 9 detection systems (industrial screeners, protein & genome LMs), Omnii leads on paraphrase F1 score. Its zero-shot score already matches or beats every baseline's multi-shot score. Less supervision, more signal. And it's not just reading structure. We stress-tested with paraphrases built to keep function but break the predicted fold. Structure-only detectors lose their main signal. Omnii degrades the least, anchoring on sequence + annotation, and not protein folds alone. Beyond single proteins, a virus is a whole system: regulatory & structural RNA set replication, host range, transmissibility. We wanted to see if Omnii could predict viral fitness effects, a proxy for host risk, on a dengue (DENV-2) deep mutational scan. Omnii predicts fitness effects from sequence alone, sans alignment. Omnii even shows structure-awareness: its strongest signal is on medium-range interactions (100-1024 nt) that fold a flavivirus on itself. Omnii reads a variant's structural environment, not just its neighbors. If models become function-aware, our defenses must too. To scale biodefense, we're scaling our team + partnerships. 🌉🌎🗼Join our mission in SF/Tokyo: https://lnkd.in/gMa7anFQ  📝 Sign up for early access: https://lnkd.in/gewyGyHk

Leaders of OpenAI, Anthropic, & GDM signed a letter urging Congress to mandate DNA synthesis screening. But chatbots & agents can't read DNA. So we built Omnii to defend against natural & AI-designed pathogens. 🖥️ Blog: https://lnkd.in/gx9stCYb 📃 Letter: screendna.org Today anyone can design DNA & get it shipped like an Amazon order from synthesis companies. This powers real science, but it's a biosecurity risk if bad actors design threats. AI raises the stakes: our prior model, Evo, generated the first AI-designed phage genome Our frontier genome language model, Omnii, reads and writes DNA/RNA and also understands proteins. It's function-aware: it screens whole genomes, not just genes/proteins, & flags threats with dangerous function even when unlike known pathogens. Ergo: generalized detection. Today's screening asks: "have we seen this before?" — aligning a sequence against known-threat databases. But sequence ≠ function. Distant proteins can fold alike, like dengue & chikungunya glycoproteins. Alignment misses this, and so we must screen beyond proteins. The hard cases are "paraphrases": same dangerous function, different sequence. Evolution designs them; a generative model designs them on demand. The kicker: a 442-residue protein can have ~10¹⁵⁶ viable variants. Enumeration is intractable. It's the blind spot of current tools. Omnii closes the gap: pathogenic functional signals live in its embedding geometry. Paraphrased pathogens cluster with their pathogenic templates, not harmless lookalikes, even when no alignment tool flags the link. A linear probe separates the classes at AUROC 0.991. Across 9 detection systems (industrial screeners, protein & genome LMs), Omnii leads on paraphrase F1 score. Its zero-shot score already matches or beats every baseline's multi-shot score. Less supervision, more signal. And it's not just reading structure. We stress-tested with paraphrases built to keep function but break the predicted fold. Structure-only detectors lose their main signal. Omnii degrades the least, anchoring on sequence + annotation, and not protein folds alone. Beyond single proteins, a virus is a whole system: regulatory & structural RNA set replication, host range, transmissibility. We wanted to see if Omnii could predict viral fitness effects, a proxy for host risk, on a dengue (DENV-2) deep mutational scan. Omnii predicts fitness effects from sequence alone, sans alignment. Omnii even shows structure-awareness: its strongest signal is on medium-range interactions (100-1024 nt) that fold a flavivirus on itself. Omnii reads a variant's structural environment, not just its neighbors. If models become function-aware, our defenses must too. To scale biodefense, we're scaling our team + partnerships. 🌉🌎🗼Join our mission in SF/Tokyo: https://lnkd.in/gMa7anFQ  📝 Sign up for early access: https://lnkd.in/gewyGyHk

Today, we’re announcing Radical Numerics’ $50 million seed round to build general biological intelligence. We’re also previewing our next-generation genome language model (gLM) called Omnii. https://lnkd.in/gPMTz7av

Scaling scientific world models requires co-designing architectures, training objectives, and numerics. Today, we share the first posts in our series on low-precision pretraining, starting with NVIDIA's NVFP4 recipe for stable 4-bit training. Part 1: https://lnkd.in/gME-VYuY Part 2: https://lnkd.in/gQWhy8G3 We cover floating point fundamentals, heuristics, custom CUDA kernels, and stabilization techniques. Future entries will cover custom recipes and results on hybrid architectures.

Sliding window attention (SWA) is powering frontier hybrid models for efficiency. Is there something better? Introducing Phalanx, a faster and better quality drop-in replacement for sliding window attention (SWA). Phalanx is a new family of hardware and numerics-aware windowed layers designed with a focus on data locality and jagged, block-aligned windows that map directly to GPUs. In training, Phalanx delivers 10–40% higher end-to-end throughput at 4K–32K context lengths over optimized SWA-hybrids and Transformers by reducing costly inter-warp communication. Today, we are releasing both the technical report, a blog, and Phalanx kernels in spear, our research kernel library. We are hiring. Blog: https://lnkd.in/gSC-H4FF Code: https://lnkd.in/grkQ2FUb Report: https://lnkd.in/guEwMsMP

Introducing RND1, the most powerful base diffusion language model (DLM) to date. RND1 (Radical Numerics Diffusion) is an experimental DLM with 30B params (3B active) with a sparse MoE architecture. We are making it open source, releasing weights, training details, and code to catalyze further research on DLM inference and post-training. We are researchers and engineers (DeepMind, Meta, Liquid, Stanford) building the engine for recursive self-improvement (RSI) — and using it to accelerate our own work.  Our goal is to let AI design AI. We’re hiring. More on RND1 models and how we are training them: - Blog: https://lnkd.in/gBWEPC5p - Code: https://lnkd.in/gemEQCvd - Report: https://lnkd.in/gbxE8FrG - Hugging Face weights: https://lnkd.in/g4gXG5Q2