Genesis-Science-1
AI agent · Research agent · Research by Arcee AI
An American open-weight model and governed research system for scientific computing workflows.
Genesis-Science-1 (GS1) is an American open-weight AI model and governed research system developed by Arcee AI in partnership with the U.S. Department of Energy. It is designed for scientific institutions and national laboratories to automate complex scientific computing workflows, such as high-performance-computing code modernization and materials science research. The system operates through a governed execution environment that maintains a reproducible record of all tool calls, code changes, and intermediate artifacts. GS1 is built to handle long-horizon tasks, including simulation campaign
- Languages
- English
- Developer
- Arcee AI
- Partner
- U.S. Department of Energy (DOE)
- Deployment
- On-premises or private infrastructure
- Context Length
- 512k (extended)
- Architecture
- Sparse Mixture-of-Experts (MoE)
Overview
Genesis-Science-1 (GS1) is an American open-weight AI model and governed research system developed by Arcee AI in partnership with the U.S. Department of Energy. It is designed for scientific institutions and national laboratories to automate complex scientific computing workflows, such as high-performance-computing code modernization and materials science research. The system operates through a governed execution environment that maintains a reproducible record of all tool calls, code changes, and intermediate artifacts. GS1 is built to handle long-horizon tasks, including simulation campaigns and experimental analysis, while ensuring human oversight for critical decisions. It provides a permanent, portable foundation that can be run on-premises within an institution's own infrastructure.
Scientific Research Capabilities
- HPC Code Modernization: It is capable of modernizing aging Fortran codebases and managing complex simulation campaigns.
- End-to-End Task Execution: The agent can carry a workflow from initial planning through to the final report.
- State Management and Recovery: It maintains task state, checkpoints progress, and manages retries and recovery during long-running tasks.
Governance and Control
- Governed Execution Environment: The model operates through a governed execution system where approved tools run in sandboxed, staged environments.
- Reproducible Audit Logs: It records all prompts, tool calls, code changes, datasets, and intermediate artifacts associated with each run.
- Human-in-the-Loop Controls: Human review is required for decisions involving safety, security, publication, and resource use.
- Controlled Access: The system does not receive blanket access to DOE systems and operates under strict permissions.
Technical Environment
- Programming Languages: Training environments include Python, Fortran, C, C++, MPI, OpenMP, CUDA, and HIP.
- Scientific Tooling: The agent works with command-line tools, notebooks, simulation packages, and computing schedulers.
- Data Integration: It utilizes experimental and observational data from user facilities and simulation outputs from supercomputing campaigns.
Specifications
- Developer
- Arcee AI
- Partner
- U.S. Department of Energy (DOE)
- Deployment
- On-premises or private infrastructure
- Context Length
- 512k (extended)
- Architecture
- Sparse Mixture-of-Experts (MoE)
Added by AntSpace from public sources (arcee.ai).