Shanghai AI Lab's Atria Dawn Matches GPT-5.6/Claude Opus 5 in Benchmarks, Excels in Real-World Tasks

Model Release and Benchmark Performance

Shanghai AI Lab has officially open-sourced Atria Dawn (书生·星河), the production version of its large language model. According to the lab's reported results across eight standard benchmarks, Atria Dawn reaches parity with Claude Opus 5 and GPT-5.6 sol, surpassing them on five of the eight benchmarks. While a gap remains to the very latest models such as GPT Astra and Opus 5.5, Atria Dawn is positioned among the top-performing open-source models.

Protein Structure Prediction Without Internet Search

In a demo conducted with internet search disabled, Atria Dawn adapted the OpenFold framework to perform protein structure prediction. Using approximately 19,000 protein chains and 2 GPUs over 3,000 training steps, the model improved local structural consistency by roughly 2.7% compared with a baseline fine-tuning configuration. On several complex selected protein cases, the predicted structures closely matched experimental structures.

CAD Modeling From Natural Language

Atria Dawn translates a single-sentence design requirement into a complete 3D CAD model with explicit part connections and motion relationships. Demonstrated examples include a rocket, motorcycle, robotic dexterous hand, and locomotive — each exhibiting proper mechanical joints rather than mere geometric stacking.

Historical Battle Visualization From Text

The model reads the original Chinese text of Romance of the Three Kingdoms , extracts entities (figures, locations, events, troop movements), and automatically generates a sandbox animation of the Battle of Guandu. Continuous narrative is converted into structured spatiotemporal relationships and then into observable movements and confrontations on the sandbox, reducing the cost of creating educational or popular-science content.

PowerPoint-Native Short Film Generation

Starting from a narrative outline, Atria Dawn produces a fully auto-playing short film entirely within PowerPoint. The demo shows 440 particles progressing through signal, waveform, spectrum, blueprint, printing, disintegration, and recombination, culminating in the Atria logo. The 29-slide deck uses only PowerPoint's native "Morph" transition; every particle and caption remains an editable native shape or text box, enabling post-generation manual edits without re-invoking the model — a step toward genuine workflow integration.

Focus on Complex Task Execution and Stability

The official release notes emphasize enhanced complex-task handling and execution stability, specifically improving requirement understanding, tool use, and execution-feedback processing. For multi-step work such as code debugging and experiment runs, Atria Dawn advances toward a goal, checks intermediate results, and adjusts subsequent actions based on feedback.