Radical Numerics Inc.
Research lab focused on large-scale AI methods and systems for biological sequences and multimodal biological data. Activities include design and pretraining of genome-focused language models, development of diffusion-based and autoregressive-to-diffusion training recipes, hardware-aligned local sequence-mixing layers and long-context modeling techniques, and systems engineering for low-precision and mixed-precision training. Outputs include technical reports, open research releases (model weights and code), and early-access collaborations for biosurveillance and sequence-level risk scoring.
Industries
Nr. of Employees
small (1-50)
Products
Research release: masked-diffusion language model and training recipe
Research release of a masked-diffusion language model (sparse MoE, research-scale parameters) accompanied by a training recipe, inference code and model weights to support reproducible experimentation in diffusion-based language modeling.
Research preview: genome language model for sequence-to-function tasks
Research-preview genome language model with long-context capability and multimodal inputs designed to support sequence-to-function prediction, variant fitness scoring and paraphrase-aware sequence detection for trusted partners.
Research release: masked-diffusion language model and training recipe
Research release of a masked-diffusion language model (sparse MoE, research-scale parameters) accompanied by a training recipe, inference code and model weights to support reproducible experimentation in diffusion-based language modeling.
Research preview: genome language model for sequence-to-function tasks
Research-preview genome language model with long-context capability and multimodal inputs designed to support sequence-to-function prediction, variant fitness scoring and paraphrase-aware sequence detection for trusted partners.
Services
Collaborative pilot partnerships providing early access to genome-model capabilities, integration of embedding-based detection into surveillance workflows, and technical support for pilot deployments.
Collaborative pilot partnerships providing early access to genome-model capabilities, integration of embedding-based detection into surveillance workflows, and technical support for pilot deployments.
Expertise Areas
- Genome language-model pretraining
- Diffusion-based language modeling and AR→diffusion conversion
- Scaling sparse mixture-of-experts architectures
- Long-context and multimodal sequence modeling
Key Technologies
- Sliding-window recurrence algorithms
- Block two-pass sequence mixing
- Hierarchical block decomposition for recurrences
- Masked-diffusion language models
News & Updates
Announcement of $50M seed financing and preview of next-generation genome language modeling research.
Research preview describing embedding-based sequence detection and organism-level fitness scoring for biodefense and surveillance applications.
Introduction of a large masked-diffusion language model trained using autoregressive-to-diffusion conversion; release of weights, code and training recipe for reproducible research.
Technical report presenting sliding-window recurrences, hierarchical decompositions, block two-pass algorithm and GPU implementation strategies for high-throughput long-context modeling.
Systems-level description of low-precision training recipes, quantization modules, tensor management and fused kernels for block-scaled GEMMs and mixed-precision linear layers.
Embedded talk on generative genomics and risks and implications of AI-designed biological sequences.
Announcement of $50M seed financing and preview of next-generation genome language modeling research.
Research preview describing embedding-based sequence detection and organism-level fitness scoring for biodefense and surveillance applications.
Introduction of a large masked-diffusion language model trained using autoregressive-to-diffusion conversion; release of weights, code and training recipe for reproducible research.
Technical report presenting sliding-window recurrences, hierarchical decompositions, block two-pass algorithm and GPU implementation strategies for high-throughput long-context modeling.
Systems-level description of low-precision training recipes, quantization modules, tensor management and fused kernels for block-scaled GEMMs and mixed-precision linear layers.
Embedded talk on generative genomics and risks and implications of AI-designed biological sequences.