Atlas

Models

← All models

Atria Dawn Preview

Shanghai AI Laboratory · 753.3B parameters

Shanghai AI Laboratory's text-only agentic instruction checkpoint built on GLM-5.2, released with MIT weights for research, coding, tool use and productivity workflows. The complete checkpoint contains 753,329,940,480 parameters; the card's 744B figure describes the foundation model. The documented deployment context is 256,000 tokens despite a larger positional limit in the compatibility config, and reasoning is configurable.

Benchmark scores

BenchmarkScore
AutomationBench 1.0.6 (provider-reported)53.8
BFCL v4 Overall Accuracy77.0
BrowseComp92.5
Capability161.0
CyberGym86.5
DeepResearch Bench II51.1
DeepSearchQA (F1 score)96.0
GDPval - AgentCompass / GLM-5.3 Elo1583
JobBench Main50.3
MLE-Bench Lite - HumanRank86.2
SkillsBench - 79 text tasks (AgentCompass)66.4
SWE-Bench Pro (Public Dataset)59.6
Terminal-Bench 2.178.3
WideSearch (reported metric)81.9
Workspace-Bench 1.0 - English Full65.0
Workspace-Bench 1.0 - English Lite68.2
τ³ Banking Knowledge - all tools41.2
Loading Atlas data…