LOGOS (arXiv 2026)

Pretrained time series foundation model for long-horizon forecasting across diverse scientific domains including climate variables, biomedical signals, and physical observations; decoder-only Transformer architecture with strong zero-shot generalization (19.8K+ stars, Apache 2.0, 2024-2025)

Foundation model for tabular data that predicts on unseen real-world tables in a single forward pass, achieving accurate small-data classification and regression without task-specific training; widely applicable to scientific datasets with limited samples (7.4K+ stars, 2022-2026)

Pretrained time series foundation model for zero-shot forecasting across diverse scientific and real-world domains; tokenizes continuous time series into discrete bins to train transformer language models on large-scale corpora, achieving strong zero-shot generalization and competitive performance with task-specific supervised models on climate, energy, and health benchmarks (5.3K+ stars, Apache 2.0, 2024-2026)

Large language model for science

Open language model for mathematics (7B/34B) trained on Proof-Pile-2, outperforming Minerva at equal scale on MATH benchmark, with tool use and formal theorem proving in Lean without finetuning (EleutherAI, ICLR 2024)

Open-source scientific multimodal foundation model built on a 235B MoE LLM and 6B vision encoder, continually pretrained on 5T tokens including 2.5T scientific-domain tokens, with strong results across chemistry, materials, life science, and earth science benchmarks (2025)