Atria-Dawn-Preview: 744B MoE-based: From research questions to verified results

This is a 744B parameter MoE-based agent model released by Shanghai AI Lab. It is designed to go beyond simple answer generation, aiming for environment understanding, tool use, and multi-step task completion.

It transforms research questions into experimental plans and automates the entire process, from code implementation to result analysis and failure recovery. It is specialized for practical tasks such as software development, data visualization, and security vulnerability verification.

It records high performance compared to competing models on search and security benchmarks, including 96.0 on DeepSearchQA and 86.5 on CyberGym. It supports a 256K context window and offers an Instruct model that processes text input only, as well as an FP8 quantized version.

It is compatible with major inference frameworks such as SGLang and vLLM. It can be integrated with coding agent environments like Codex and Claude Code to perform complex engineering tasks.