From scientific research to 3D creation, Atria Dawn wants AI to work on end-to-end projects

Atria Dawn is exiting its Preview phase with 512,000 context tokens, enhanced agentic capabilities, and demonstrations in research, 3D, and creation.

Training a protein structure prediction model, drawing a 3D robotic hand, or producing a fully editable animation in PowerPoint. These projects illustrate the ambitions of Atria Dawn, the artificial intelligence model developed by the Shanghai Artificial Intelligence Laboratory. Its official version succeeds Atria Dawn Preview with an expanded context window of 512,000 tokens, enhanced tool-use capabilities, and new results on long-horizon professional tasks. A Model Built to See Projects Through

Atria Dawn belongs to the category of high-capacity agentic models, featuring an architecture of approximately 700 billion parameters resulting from the Chinese laboratory's work.

Its distinctiveness lies in how it approaches complex tasks. Rather than merely producing answers or code snippets, it seeks to drive a project through multiple stages: understanding objectives, using tools, execution, verifying results, and correcting errors.

The Preview version already featured a 256,000-token context window. Doubling it now allows the model to retain more information during extended tasks, particularly when an agent must consult numerous files, track experiments, or revisit previous decisions.

Improvements also extend to machine learning programming, with particular attention paid to the stability of processes requiring extensive interactions with tools. From Scientific Experiments to 3D Models

One of the most technical demonstrations involves protein structure prediction. In an environment deprived of web access, Atria Dawn used OpenFold, a framework based on Google DeepMind's AlphaFold 2 architecture, to conduct experiments and optimize an existing system.

The model participated in fine-tuning over 80 million parameters across approximately 19,000 protein chains, during 3,000 main training steps using two GPUs.

According to the laboratory, the optimized setup improves local structural consistency by approximately 2.7% compared to the baseline setup during validation on the same dataset. This result measures a targeted experimental improvement, rather than a general superiority over specialized protein prediction systems.

The demonstrations also cover computer-aided design. From textual descriptions, Atria Dawn can produce geometric models and technical drawings representing rockets, motorcycles, locomotives, or articulated robotic hands, among others. When PowerPoint Becomes an Animation Tool

Another demonstration explores creation possibilities directly within PowerPoint, without relying on video editing software.

Based on a script, Atria Dawn designed an animation using approximately 440 graphic elements and 29 slides. Objects move, scatter, and reassemble using the software's native Morph transitions.

Every element, including text, remains editable within the presentation. The result is therefore not an imported video, but an animated composition built using PowerPoint's own tools.

The model can also analyze a scientific paper to extract its objectives, methods, and results, then organize them into an academic poster featuring text, graphics, and equations.

In a different vein, it transforms a passage from the historical novel Romance of the Three Kingdoms into an animated map of the Battle of Guandu, complete with a representation of military movements and their evolution.

These examples highlight the value of an AI capable of producing usable and editable assets, beyond simple image generation. Progressing Results, but Uneven Across Tasks

Evaluations published with the official version show several advances compared to Atria Dawn Preview.

On AutomationBench, dedicated to executing automated tasks, the score increases from 53.8 to 54.7. On Workspace-Bench, it progresses from 65 to 66.1, while MLE-bench Lite reaches 88.2, compared to 86.2 previously.

On WideSearch, dedicated to information retrieval across multiple sources, the official chart displays 84.6, compared to 81.9 for the Preview. However, the announcement published on 𝕏 mentions 84.8, a slight discrepancy between the two mediums.

Performance does not improve across the board. On SWE-bench Pro, Atria Dawn drops from 59.6 to 59.4, while certain competitors maintain a significant lead in software development.

On JobBench, however, the model advances markedly, from 50.3 to 57.5. It notably surpasses GPT-5.6 Sol in the published benchmark, though it remains behind Claude Opus 5.

These results stem primarily from evaluations compiled by the laboratory, using different environments and protocols. They do not constitute a universal measurement of the model's reliability. Deliverables Rather Than Simple Answers

Atria Dawn also explores professional data analysis. In one demonstration, the model reconciles five sales files, calculates profits across different product categories, identifies anomalies, and generates visualizations comparing results against targets.

The laboratory also presents several creations made by its community: web applications generated from a prompt, an interactive visualization of continental drift, and a simulation of a sight-seeing flight over Hong Kong.

The official Atria website provides access to these demonstrations, while the project's resources are available on GitHub, Hugging Face, and ModelScope. The playground platform also allows users to experiment with the models according to available access methods.

With this release, the Shanghai Artificial Intelligence Laboratory defends an approach to AI where quality is also measured by what remains after execution: a working program, a documented experiment, a 3D model, or an editable presentation.

The demonstrations remain curated by the developers and do not guarantee equivalent success across all projects. They nonetheless illustrate an important direction in agentic research: shifting models from content generation to the completion of full-scale work, whose results can be examined, corrected, and reused.