Atria Dawn Preview targets long-running agent workloads

Shanghai AI Laboratory has released Atria Dawn Preview, a 744-billion-parameter Mixture-of-Experts foundation model designed to execute multi-step tasks with tools and environmental feedback. The project code and published weights carry the MIT License, giving developers broad rights to inspect, modify, host, and commercialize the model.

The preview includes a standard instruction-tuned checkpoint and an FP8 variant intended to reduce memory use and potentially improve inference throughput on compatible hardware. Its scale, permissive terms, and emphasis on complete task trajectories expand the options for teams building self-hosted research, automation, and security agents.

Built for long task runs

Atria Dawn Preview is designed to maintain a plan while tools return new information. Its proposed workflow covers problem analysis, solution design, tool selection, code generation, experiment execution, result analysis, and recovery from failed steps. The team groups its intended workloads into four categories:

• Discovery: Retrieve and organize evidence, conduct deep research, and convert research questions into executable experimental plans.
• Creation: Build software, interactive applications, games, data visualizations, and machine-learning systems.
• Delivery: Convert documents, data, and design requirements into reports, presentations, and other structured outputs.
• Cybersecurity: Analyze security issues, validate vulnerabilities, draft fixes, and retest systems in authorized environments.

Cybersecurity receives unusual prominence for a general-purpose model release. Shanghai AI Lab reports a CyberGym score of 86.5, the highest result in its published comparison table. That claim applies to the release team's evaluation setup, including its tools, prompts, budgets, and scoring rules.