Modular

Small, efficient, honestly benchmarked models for science and for languages the big labs skip.

Modular is an AI research and data-infrastructure company founded in 2026 and based in Nigeria. We build compact open models, the datasets and pipelines behind them, and governed data infrastructure for biomedical research. Our view is that a lot of useful AI doesn't need frontier-scale compute: a small model that is trained carefully and evaluated honestly can run on a CPU, cost very little, and still compete.

🌐 modularcomputing.cloud · ✉️ contact@modularcomputing.cloud

Research interests

• AI for biology and chemistry. This covers protein stability and mutation-effect prediction, small-molecule identification from tandem mass spectrometry (MS/MS), de novo spectrum-to-structure generation, and single-cell annotation.
• Speech and language for Nigerian languages. We work on natural code-switched text-to-speech, speech recognition and full-duplex voice agents for Igbo, Yoruba, Hausa, Nigerian Pidgin and English, plus cross-lingual embeddings and image-text retrieval.
• Efficient foundation models. This includes compact time-series forecasting, compute-efficient video-to-3D reconstruction, and vision-language models for local languages.
• Biomedical data infrastructure. With the Institute of Genomics and Global Health (IGH) at Redeemer's University and its CAMRA antimicrobial-resistance programme, we are building consented, governed pipelines that collect and curate clinical laboratory data.

Models

Area	Model	What it does
Protein	StaMP	Predicts single-mutation protein stability (ΔΔG) with a small head on frozen ProteinMPNN features. Spearman 0.734 on the Megascale test split. Full saturation scans run on a CPU. MIT licence.
Speech	dynamic-accent-pro-v1.1	Text-to-speech in English, Pidgin, Yoruba, Igbo and Hausa that switches language mid-sentence in one voice.
Speech	dynamic-asr	Code-switched speech recognition for the same five languages.
Speech	dynamic-accent-duplex	A full-duplex speech-to-speech voice agent.
Language	Native-Bird	Cross-lingual sentence embeddings for Igbo, Hausa, Yoruba and English.
Multimodal	Native-Bird2	Image-text retrieval for Nigerian languages.
Multimodal	AtlasVision	A vision-language model built on N-ATLaS and SigLIP2.
Time series	Modular-TS-2.4M	A 2.4M-parameter zero-shot probabilistic forecasting model.
3D	Polymorphic	Experiments in compute-efficient video-to-3D reconstruction.

Also in development: an MS/MS molecular-identification pipeline that ranks about 1M candidate structures, a de novo spectrum-to-structure model for molecules outside existing databases, and C2S, an LLM-based cell-type annotation tool grounded in Gene Ontology.

How we work

• Small and efficient first. We aim for models that are cheap to train, cheap to run, and usable without a GPU cluster.
• Honest evaluation. Every model card reports out-of-domain results and known weaknesses next to the headline score.
• Open by default. Code and weights are released under permissive or share-alike licences wherever the training data allows it.

Interested in collaborating, piloting a model, or partnering on data? Write to contact@modularcomputing.cloud.