I Ran Atria Dawn for 72 HOURS (China's New #1 Local AI?)

Summary

The video reviews Atria Dawn—a model name inconsistently transcribed in the subtitles—as a preview-stage model from Shanghai AI Laboratory. The presenter says it is built on GLM 5.2 and claims its benchmark rankings place it ahead of models including Qwen, Opus, and GLM. Rather than treating those rankings as conclusive, the video tests the model through local coding and application-generation tasks.

Local setup and inference

The model was run across a 128GB M4 Max MacBook Pro and a 512GB Mac Studio. After finding the output unreliable at 4-bit quantization, the presenter increased it to roughly 5.5-bit. Generated code still often required debugging, and the presenter says the fully unquantized model was beyond the available hardware.

The model offers reflection, or “thinking,” modes. These sometimes generated tens of thousands of tokens—and in some cases more than 100,000—making generations take hours. The presenter recommends setting a limit, such as 10,000–25,000 thinking tokens, rather than letting reasoning run indefinitely.

Coding and interactive-generation tests

The presenter tested the model by generating 3D scenes, games, and software interfaces. Runtime and compilation errors were common, and many demos worked only after the presenter fixed or changed the code. Even so, some results were functional or visually impressive.

3D scenes and visual demos

• Solar systems: The model produced attractive browser-based scenes, though shader and compilation errors occurred. A Python/Matplotlib solar system worked with relatively little output. A more elaborate spaceship scene ran but had obstructive interface elements.
• Scientific and visual demos: A nuclear-explosion simulation was underwhelming compared with the presenter’s GLM 5.3 example. A procedural human face became a standout after fixes, with adjustable lighting, rendering quality, rotation, and facial expression. An attempt at 3D anatomy did not complete.

Games

• Call of Duty- and GTA-style demos had problems such as dark visuals, weak gameplay, and missing or faulty collisions.
• A Red Dead Redemption-style scene had a polished menu and convincing desert setting, but confusing, broken gameplay.
• An amusement-park scene appeared to show its minimap instead of the intended 3D view. The flight simulator looked promising, but the player or camera was misplaced.
• A kart racer had a backward-facing camera and no collisions.
• A Super Mario-style platformer had an appealing level and enemies, but movement and runtime problems required fixes.
• A Final Fantasy VII scene included recognizable characters, but had poor lighting, broken collisions, and little interactivity.
• Street Fighter II was one of the more entertaining results. The characters could fight, though the presentation and moves were rough.
• A Tomb Raider-style level included moving hazards, collectible objectives, and a rolling-ball sequence, but the character lacked a head.
• An Age of Empires II-style game was a notable success after fixes: the presenter could gather resources, create units, build structures, and attack an opponent.

Productivity and creative apps

• A Word clone had a basic interface despite an exceptionally long generation and a runtime error.
• A PowerPoint-style app supported slide navigation, presentation mode, and text editing, though it also encountered errors.
• A Photoshop-style demo allowed some drawing, while an iMovie-style timeline showed a simulated transition.
• Logic Pro- and Blender-style interfaces showed music-timeline and object-manipulation features, respectively. However, the Blender demo did not produce the expected rendered output.

Overall assessment

The presenter’s verdict is mixed. Atria Dawn showed notable ability to generate interactive applications and games, and several results—especially the procedural face, Street Fighter-style demo, and Age of Empires-style game—had clear potential. However, the preview was inconsistent: frequent execution errors, excessive output, weak lighting, and broken gameplay often meant substantial repair or additional prompting. The presenter also says basic logic tests produced results similar to GLM 5.3.

The model is described as having a commercially friendly MIT license, with the caveat that it was still a preliminary release.

Main speakers and sources

• Speaker: The video’s host from xCreate, who conducted the local tests and evaluation.
• Sources discussed: Shanghai AI Laboratory as the model developer; ModelScope as the model-hosting site; and GLM 5.2/5.3 and Qwen as comparison points.