Shanghai AI Laboratory logo

Atria Dawn Preview

Shanghai AI Laboratory

Atria Dawn Preview is a preview-stage agentic LLM from Shanghai AI Laboratory (released under the internlm Hugging Face organization), built on a 744B-parameter MoE foundation that the model card says was trained from the GLM-5.2 base. It targets long-horizon research and engineering workflows, with the vendor pitching four capability areas: Discovery, Creation, Delivery and Cybersecurity. The instruct checkpoint is text-only with a 256K-token context window, ships in BF16 and FP8 under the MIT license, and is served through the international api.atria-asi.ai endpoint (Chat Completions, Messages and Responses formats) with no published per-token pricing.

Key Specifications

Parameters
744.0B
Context
-
Release Date
September 11, 2026
Average Score
72.7%

Timeline

Key dates in the model's history
Announcement
September 11, 2026
Last Update
September 16, 2026
Today
September 20, 2026

Technical Specifications

Parameters
744.0B
Training Tokens
-
Knowledge Cutoff
-
Family
-
Capabilities
MultimodalZeroEval

Benchmark Results

Model performance metrics across various tests and benchmarks

Other Tests

Specialized benchmarks
AutomationBench
AutomationBench. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
53.8%
BFCL-V4
BFCL v4. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
77.0%
BrowseComp
BrowseComp. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
92.5%
CyberGym
CyberGym. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
86.5%
DeepSearchQA
DeepSearchQA. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
96.0%
GDPval-AA
GDPval row (Elo-style scale, e.g. Fable 5.1 reports 1853 on the same scale). Mapped to gdpval-aa (max_score 3000) rather than the fractional gdpval benchmark because the reported value is on the Elo index scale, not a 0-1 win rate. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
52.8%
MLE-Bench Lite
MLE-bench Lite. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
86.2%
SkillsBench
SkillsBench. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
66.4%
SWE-Bench Pro
SWE-bench Pro. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
59.6%
Tau3 Banking
τ³-Bench Banking. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
41.2%
Terminal-Bench 2.1
Terminal-Bench 2.1. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
78.3%
WideSearch
WideSearch. Official Atria Dawn Preview Hugging Face model card evaluation table.Self-reported
81.9%

License & Metadata

License
mit
Announcement Date
September 11, 2026
Last Updated
September 16, 2026

Compare Atria Dawn Preview

All comparisons

Similar Models

All Models

Recommendations are based on similarity of characteristics: developer organization, multimodality, parameter size, and benchmark performance. Choose a model to compare or go to the full catalog to browse all available AI models.