Find open-source science resources

Benchmark evaluating AI agents on 75 curated Kaggle-style ML engineering competitions with reproducible Docker-based grading harness, human baselines, and end-to-end task lifecycle, used as a primary benchmark for autonomous ML research agents (e.g., InternAgent #1 at 36.44%)

For a convenient overview and download list, visit our model page for this model.

Protein structure prediction

Distributional flow matching model for robust single-cell perturbation prediction, modeling the full distribution of perturbed cellular expression profiles conditioned on control states via PAD-Transformer and multi-kernel MMD regularization; reduces MSE by 19.6% over the strongest baseline in combinatorial settings (Westlake University, 41+ stars, MIT License)

FutureHouse's end-to-end scientific discovery multi-agent system orchestrating literature search (Crow/Falcon) and data analysis (Finch) agents, first AI-generated drug discovery identifying ripasudil as novel dry AMD therapeutic (2025)

Benchmark evaluating AI agents' ability to replicate 20 ICML 2024 Spotlight/Oral papers from scratch, with 8,316 gradable tasks and author-co-developed rubrics

End-to-end molecular dynamics engine built on PyTorch, enabling differentiable simulations with neural network potentials and GPU acceleration for machine learning-accelerated molecular dynamics (MIT License, 707+ stars)

A 50% magnitude-pruned version of facebook/esm2t1235MUR50D optimized for efficient drug discovery inference on Apple Silicon.

Scaling efficient, expressive, and general SE(3)-equivariant graph attention transformers for atomic systems and machine-learned interatomic potentials (MIT License, 2026)

Pretrained machine-learned force field for (bio)molecular simulations combining the fast SO3krates neural network for semi-local interactions with universal pairwise force fields for short-range repulsion, long-range electrostatics, and dispersion interactions; supports geometry optimization, NVT/NPT/NVE MD, fine-tuning, ASE calculator, and JAX-MD integration (JACS 2025, 218+ stars, MIT License)

Duchifat-2.3-Instruct is a state-of-the-art, instruction-tuned Large Language Model developed by TopAI. As the flagship of the Duchifat series, this model represents a fundamental breakthrough in how Hebrew is processed, reasoned, and generated in the LLM era.

!image

For a convenient overview and download list, visit our model page for this model.

A generalist foundation model for healthcare capable of handling diverse medical data modalities.

Fully autonomous medical image segmentation research system that generates complete manuscripts end-to-end from datasets with zero human intervention, beating strongest baselines on 24 of 31 datasets and achieving T1-T2 tier manuscript quality in double-blind evaluations (USTC & Shanghai AI Lab, 2026)

Unified latent diffusion transformer that jointly generates periodic crystals and non-periodic molecules, scaling to 500M parameters with SOTA results on QM9, MP20, and GEOM-DRUGS (Meta FAIR, ICML 2025, 310+ stars)

Self-attention transformer performing simultaneous earthquake detection and P/S phase picking on continuous seismic waveforms, trained on the large-scale STEAD benchmark dataset and outperforming legacy detection methods with far fewer false positives; widely adopted by seismological observatories for earthquake monitoring, included in SeisBench model collections, and extended by efficient EQT-Mini/EQT-Lite successors for real-time deployment (Stanford, 418+ stars, MIT License, actively maintained)

AstraZeneca's industrial-grade retrosynthetic planning tool using MCTS to recursively decompose molecules into purchasable precursors, with multi-step route scoring and support for custom one-step models (v4.0, 2024)

Fine-tuned version of google/gemma-4-E4B-it across three professional domains — Medical, Legal, and Finance — using QLoRA (4-bit NF4) with Optuna-tuned hyperparameters, trained on Kaggle T4 GPU.

Scalable agentic training environment for code-centric reasoning in biomedical data science

L1 (Learning Unit 1) is the first language model from Lunit and Lunit Consortium, purpose-built for the medical domain. Derived from Gravity-16B-A3B-Base, L1 is designed for clinical reasoning and decision support.

Improved equivariant Transformer for 3D atomic graphs (ICLR2024)

DeepConsensus uses gap-aware sequence transformers to correct errors in Pacific Biosciences (PacBio) Circular Consensus Sequencing (CCS) data.

Equivariant foundation models for materials "on a budget", releasing pre-trained universal interatomic potentials (nequix-mp-1 on MPtrj, nequix-omat-1 on OMat24, nequix-oam-1 combining OMat24, sAlex, and MPtrj frontier datasets) with phonon fine-tuning (PFT) for accurate lattice dynamics and analytical Hessians; pip-installable with ASE calculator and JAX/PyTorch backends with OpenEquivariance kernels (Atomic Architects, 76+ stars, MIT License, 2025-2026)

Extended autonomy AI scientist with 200 parallel agent rollouts, 42K lines of code execution, 1.5K papers analyzed per run, achieving 79.4% accuracy and 7 scientific discoveries (Edison Scientific)

## Model Description This is a lightweight, high-performance image classification model built to diagnose histopathological scans of lung and colon tissues. This model was specifically designed for rapid web deployment without sacrificing clinical accuracy.

LLM agent framework for Earth Observation with 104 specialized tools across 5 functional kits

Tool to build force field input files for molecular simulation.

GFF and GTF file manipulation and interconversion.

Baidu's open-source reproduction of AlphaFold3 in PaddlePaddle, providing pretrained weights and inference pipelines for unified biomolecular structure prediction across proteins, nucleic acids, ligands, ions, and post-translational modifications within the PaddleHelix biocomputing platform (Baidu, bioRxiv 2024)

Hybrid deep learning and alignment-based tool for identifying viruses, plasmids, and other mobile genetic elements in isolates, metagenomes, and metatranscriptomes, combining neural-network gene-content classifiers with nucleotide-sequence signatures; also performs viral taxonomic assignment, provirus detection in host genomes, and functional annotation, with precomputed databases of 200K+ viral and 1M+ plasmid genomes and web apps on Galaxy and NMDC EDGE (Berkeley Lab & DOE Joint Genome Institute, 334+ stars, actively maintained)

Allen Institute for AI's global geospatial foundation model for satellite imagery analysis, enabling large-scale mapping of buildings, wind turbines, trees, and land cover from Sentinel-2 data with open-source weights and inference tools (2024)

First agentic framework for weather science, pairing an LLM with ZephyrusWorld (a code-execution environment exposing WeatherBench 2 data, geolocation, forecasting, simulation, and climatology tools) and ZephyrusBench (2,230 Q&A pairs across 49 weather-science tasks); outperforms text-only baselines by up to 44.2 percentage points (UC San Diego Rose-STL-Lab, 99+ stars, MIT License, 2026)

Unified ML/DL framework for drug discovery workflows, integrating RDKit, DeepChem, and scikit-learn with SHAP explainability

Andrej Karpathy's autonomous LLM research framework: AI agent runs overnight experiments on a real training setup, auto-editing code→5min training→evaluation in a loop, ~100 experiments per night on a single GPU

Toolkit for linearizing academic PDFs into LLM-ready text with high accuracy and structure preservation, optimized for scientific literature extraction

End-to-end semi-automated scientific discovery system that designs, iterates, and analyzes code-based experiments via LLM-as-a-mutator over scientific articles and code examples; auto-creates, runs, and debugs experiment code in containers and writes meta-analysis reports (339+ stars, Apache 2.0)

Automated code generation from machine learning research papers into runnable implementations (4.5K+ stars, 2025)

MarkushGrapher-2 is an end-to-end multimodal model for recognizing chemical structures from patent document images. It jointly encodes vision, text, and layout information to convert Markush structure images into machine-readable CXSMILES representations.

AlphaFold fine-tuned with flow matching for generating protein conformational ensembles, covering both experimental PDB states and molecular dynamics ensembles at physiological temperatures; includes ESMFlow variant (MIT, 526+ stars, 2024)

ChemicalOCR is a compact vision-language model fine-tuned specifically for optical character recognition (OCR) in chemical structure images. It extracts text and bounding boxes from molecular drawings, enabling the recognition of atom labels, abbreviations, and descriptive text within chemical…

Deep learning-based variant caller

Bi-directional DNA language model based on the Mamba state space architecture, enabling efficient long-range genomic sequence modeling with linear-time complexity and built-in reverse-complement equivariance; achieves strong performance on chromatin accessibility, enhancer, and promoter prediction benchmarks (Stanford & UC Berkeley, 500+ stars)

The Graphic Descriptor Ontology (GDO) is intended for use in describing graphics that represent the form of objects. It uses the language of visual communication, illustration, and technical drawing. The GDO is rooted in the Basic Formal Ontology (BFO) and uses several classes from the Information Entity Ontology of the Common Core Ontologies as a mid-level ontology. [from https://gdo.endlessforms.info/about]

Transform arXiv papers into Beamer slides using LLMs

Universal scientific research intelligence covering 50+ disciplines, repositioning LLMs as cross-disciplinary generators with human experts as verifiers; 30B model outperforms Claude Opus and GPT on 5 research benchmarks

# GigaHeart ## A Cardiac-specific CT Foundation Model for Heart Transplantation

![Language: Multilingual]()