Shanghai AI Lab introduces Atria Dawn, a 744B-parameter agentic AI model built on GLM-5.2

Shanghai Artificial Intelligence Laboratory has introduced the full version of Atria Dawn, a large-scale agentic AI model built on the 744-billion-parameter Mixture-of-Experts architecture of GLM-5.2, targeting complex research, software engineering, office productivity and cybersecurity workflows.

The updated release, documented in October 2026 following Atria Dawn Preview in September, focuses on completing multi-stage tasks requiring continuous interaction with external environments. The model is designed to investigate problems, use tools, execute operations, inspect results and revise its approach when an attempt fails.

Shanghai AI Lab reports improvements over the Preview across JobBench, WideSearch, PaperBench, SWE-bench Pro and Terminal-Bench 2.1, alongside areas where performance remains below competing models. Full-version BF16 and FP8 downloads are still listed as forthcoming, while Preview checkpoints are available through Hugging Face and ModelScope.

Although some release descriptions mention a 512K-token context, the official model download table and published client configurations specify 256K tokens. Until clarified, 256K is the more defensible documented deployment limit.

··········

ATRIA DAWN BUILDS AN AGENTIC SYSTEM ON A 744-BILLION-PARAMETER FOUNDATION MODEL.

Atria Dawn is developed by Shanghai AI Lab using GLM-5.2 as its foundation. Mixture-of-Experts architecture distributes capacity across specialized neural-network components and activates selected experts during processing. 744 billion parameters describe total scale, not necessarily the number activated per generated token.

........

Specification	Atria Dawn
Developer	Shanghai Artificial Intelligence Laboratory
Foundation model	GLM-5.2
Total parameter scale	744 billion
Architecture	Mixture-of-Experts
Model specialization	Agentic research and engineering
Supported input modality	Text
Documented context length	256K tokens
Hosted access	International and China API services
Preview weights	Available
Full-version weights	Coming soon
Preview formats	Standard and FP8
Repository license	MIT
Deployment frameworks	SGLang and vLLM

........

The project targets discovery, creation, delivery and cybersecurity. These domains share the requirement to maintain a coherent objective across multiple tool interactions, where feedback and intermediate failures determine the next operation.

··········

THE OCTOBER RELEASE EXTENDS THE ATRIA DAWN PREVIEW INTRODUCED IN SEPTEMBER.

Atria Dawn Preview and research materials were made public in September 2026. The updated repository later presented full-version results alongside Preview evaluations. Public documentation, hosted access and downloadable weights are distinct forms of availability.

........

Release component	Documented status
Atria Dawn Preview	Published September 2026
Preview standard checkpoint	Available on Hugging Face and ModelScope
Preview FP8 checkpoint	Available
Atria Dawn full version	Documented in October materials
Full-version hosted API	Documented
Full-version standard weights	Coming soon
Full-version FP8 weights	Coming soon
Technical report	Published

........

The full version's reported benchmark improvements cannot automatically be attributed to older downloadable Preview weights. The availability distinction matters for reproducibility, deployment planning and independent evaluation.

··········

THE MODEL TARGETS COMPLETE WORKFLOWS RATHER THAN ISOLATED ANSWERS.

Atria Dawn aims to transform open-ended assignments into executable, verifiable results through problem analysis, tool selection, implementation, execution and error recovery. Producing plausible explanations alone does not demonstrate successful completion in an external environment.

A research agent may need to find evidence, write and run analysis scripts, correct errors and produce reproducible findings. A coding agent may inspect a repository, implement changes and verify that tests pass.

........

Agentic domain	Representative operations	Expected deliverable
Discovery	Search, evidence collection, planning	Research findings and executable plans
Creation	Programming and application development	Working software and artifacts
Delivery	Documents, spreadsheets and presentations	Structured business deliverables
Cybersecurity	Analysis, validation and remediation	Tested fixes and records

........

These workflows depend on feedback from tools and environments. Unexpected results must influence subsequent decisions, while completion should be measured against explicit requirements.

··········

JOBBENCH IMPROVES BY 7.2 POINTS OVER THE PREVIEW MODEL.

Shanghai AI Lab reports 57.5 on JobBench, compared with 50.3 for Preview: a 7.2-point gain, approximately 14.3% relative to the earlier score. Other benchmarks show smaller but measurable changes.

........

Benchmark	Atria Dawn	Preview	Difference
JobBench	57.5	50.3	+7.2
WideSearch	84.6	81.9	+2.7
PaperBench	89.6	87.0	+2.6
SWE-bench Pro	61.7	59.6	+2.1
Terminal-Bench 2.1	81.0	78.3	+2.7
AutomationBench	54.7	53.8	+0.9
SkillsBench	67.8	66.4	+1.4
Workspace-Bench-Lite	70.0	68.2	+1.8
Workspace-Bench	66.1	65.0	+1.1
SpreadsheetBench 2	33.0	31.5	+1.5

........

These are developer-reported results, not independent reproductions. Improvements vary substantially by task, making workload-specific evaluation more informative than a single aggregate score.

··········

SEARCH AND RESEARCH BENCHMARKS SHOW STRONG PERFORMANCE WITH IMPORTANT EXCEPTIONS.

Research agents must retrieve evidence, organize findings, resolve conflicts and decide whether additional verification is needed. Atria Dawn improves on selected search and paper benchmarks, but not every evaluation.

........

Research benchmark	Atria Dawn	Preview	Change
DeepSearchQA	96.5	96.0	+0.5
WideSearch	84.6	81.9	+2.7
DeepResearch Bench II	51.2	51.1	+0.1
BrowseComp	91.3	92.5	−1.2
PaperBench	89.6	87.0	+2.6

........

BrowseComp declines by 1.2 points, and DeepResearch Bench II barely changes. These results prevent a blanket claim of superior research performance. Retrieval scores alone cannot guarantee correct interpretation, source provenance or reproducible scientific analysis.

··········

SOFTWARE ENGINEERING RESULTS ARE COMPETITIVE BUT DO NOT ESTABLISH UNIVERSAL LEADERSHIP.

Atria Dawn targets repository inspection, implementation, terminal execution, testing and correction. SWE-bench Pro improves to 61.7 from 59.6, while Terminal-Bench 2.1 rises to 81.0 from 78.3.

........

Software engineering benchmark	Atria Dawn	Preview
SWE-bench Pro	61.7	59.6
Terminal-Bench 2.1	81.0	78.3
NL2Repo-Bench	50.0	51.4
DeepSWE	49.3	Not reported

........

NL2Repo-Bench declines, illustrating differences between terminal operation, repository reasoning and extended autonomous coding. Production evaluations should also measure test completion, unintended file modifications, tool reliability and review burden.

··········

COMPARISONS WITH GPT, CLAUDE, QWEN AND GLM REQUIRE BENCHMARK-SPECIFIC INTERPRETATION.

Shanghai AI Lab compares Atria Dawn with GPT-5.6 Sol, Claude Opus 5, Qwen 3.8 Max and other contemporary models. No single model leads across every category.

........

Benchmark	Atria Dawn	GPT-5.6 Sol	Claude Opus 5	Qwen 3.8 Max
WideSearch	84.6	83.3	84.0	81.9
PaperBench	89.6	92.8	88.9	94.1
JobBench	57.5	45.4	68.0	52.7
SWE-bench Pro	61.7	59.8	76.5	65.3
Terminal-Bench 2.1	81.0	86.2	90.2	89.6
Workspace-Bench	66.1	56.0	65.8	68.0
SpreadsheetBench 2	33.0	32.7	33.6	34.9

........

Atria Dawn leads some listed models on WideSearch and Workspace-Bench but trails Claude Opus 5 substantially on SWE-bench Pro and Terminal-Bench 2.1. Evaluation harnesses, tool permissions and compute budgets may affect comparisons. These figures are the laboratory's published table, not a uniformly independent ranking.

··········

THE 256K AND 512K CONTEXT FIGURES REQUIRE CAREFUL DISTINCTION.

Some descriptions of the full release mention 512K tokens, but the official download table specifies 256K for full and Preview variants, including FP8. Published Codex and Kimi Code configurations also specify 256,000 tokens.

........

Official documentation component	Context specification
Atria Dawn download table	256K
Atria Dawn FP8 download table	256K
Preview download table	256K
Preview FP8 download table	256K
Codex configuration	256,000 tokens
Kimi Code configuration	256,000 tokens
Some release descriptions	512K claimed

........

The difference could reflect planned capabilities or hosted settings, but the primary materials do not reconcile it sufficiently. 256K is the defensible documented deployment limit until clarified. A large context window also does not guarantee equally accurate use of all information within it.

··········

TOOL USE IS CENTRAL TO ATRIA DAWN'S AGENTIC PERFORMANCE.

Agents must select tools, format calls correctly, interpret responses and decide subsequent actions. Atria Dawn improves modestly across several tool-use evaluations.

........

Tool-use benchmark	Atria Dawn	Preview	Change
BFCL V4	78.4	77.0	+1.4
τ³-bench Banking	43.3	41.2	+2.1
SkillsBench	67.8	66.4	+1.4
AutomationBench	54.7	53.8	+0.9

........

Function-calling accuracy is different from long-horizon workflow completion. A correctly formatted call may still be unauthorized or harmful, so deployment needs permission controls, execution limits and audit logs.

··········

CYBERSECURITY CAPABILITIES REQUIRE AUTHORIZED AND VERIFIABLE EXECUTION.

Cybersecurity is one of the project's four target domains. Shanghai AI Lab describes vulnerability analysis, validation, remediation and verification. CyberGym rises from 86.5 to 86.8, while Harvey LAB is reported at 27.5 for the full model.

These scores indicate explicit security-task evaluation but not unrestricted operational reliability. Security agents should operate within authorized environments, with action logs, constrained permissions and human approval for potentially destructive operations.

Identifying a suspected vulnerability differs from proving that a remediation works without disrupting legitimate functionality. Production suitability depends on the execution environment as well as model capability.

··········

PREVIEW WEIGHTS ARE AVAILABLE WHILE FULL-VERSION DOWNLOADS REMAIN PENDING.

The official project provides Preview standard and FP8 checkpoints. Full Atria Dawn and full FP8 entries remain labeled Coming soon. Repository licensing should not be confused with immediate availability of every advertised checkpoint.

........

Deployment option	Availability
Hosted international API	Documented
Hosted China API	Documented
Preview standard checkpoint	Available
Preview FP8 checkpoint	Available
Full standard checkpoint	Not yet linked
Full FP8 checkpoint	Not yet linked
SGLang deployment guidance	Available
vLLM deployment guidance	Available

........

The laboratory documents SGLang 0.5.13.post1 or later and vLLM 0.23.0 or later. Developers should verify support for the exact checkpoint, precision and model architecture rather than assuming generic transformer compatibility.

··········

HOSTED API ACCESS OFFERS A LOWER-BARRIER DEPLOYMENT ROUTE.

Shanghai AI Lab documents international and China API services using the model identifier Atria-Dawn, including configurations for coding-agent clients through an OpenAI-compatible interface.

The documented input modality is text only. Developers should not assume native screenshot, image or PDF-page understanding; visual workflows may need preprocessing or a separate vision-capable model.

Hosted inference also involves rate limits, service availability, data handling, latency and pricing. The examined official materials do not establish a complete verified price schedule suitable for cost-per-million-token comparison, so no speculative API prices are assigned.

··········

SELF-HOSTING A 744B MODEL REQUIRES SUBSTANTIAL MEMORY AND COMPUTE.

A 744-billion-parameter model requires considerable infrastructure even when expert routing reduces active computation. Theoretical raw weight storage can be calculated directly.

........

Precision assumption	Theoretical raw weight storage
BF16, 16-bit	1.488 TB
FP8, 8-bit	744 GB
Hypothetical 4-bit	372 GB

........

These decimal-unit calculations exclude metadata, activation buffers, attention caches and serving overhead. The 4-bit case is illustrative, not confirmation of an official 4-bit release.

Mixture-of-Experts sparsity can reduce arithmetic per token but does not automatically remove the need to store or access the full expert collection. Depending on routing, offloading and hardware, deployment may require large aggregate accelerator memory and high-bandwidth interconnects. Hosted access is therefore simpler for many smaller organizations.

··········

AGENTIC PERFORMANCE SHOULD BE MEASURED BY COMPLETED WORK AND TOTAL COST.

A model may generate accurate intermediate answers while failing the overall task, or require several attempts before producing a correct deliverable. Evaluating the entire workflow is essential.

........

Operational metric	What it measures
Task completion rate	Assignments successfully finished
Deliverable validity	Satisfaction of requirements
Tool-call accuracy	Correct tool selection and execution
Recovery success	Resolution of intermediate failures
Human intervention rate	Manual assistance required
Execution time	End-to-end latency
Token and tool consumption	Total resource usage
Cost per successful task	Economic efficiency
Unauthorized-action rate	Compliance with boundaries
Reproducibility	Ability to verify and repeat results

........

A research report without reproducible calculations may be unsuitable for scientific work. A coding agent that changes unrelated files can impose review and maintenance costs. A spreadsheet without traceable formulas may not meet financial-control standards. The relevant goal is correct, auditable work at an acceptable total cost.

··········

THE RELEASE ALSO EXPOSES THE LIMITS OF CURRENT AGENT BENCHMARKS.

The updated results are mixed: JobBench improves 7.2 points, DeepResearch Bench II only 0.1, while BrowseComp and NL2Repo-Bench decline. Agentic competence is not one universal property.

Benchmark outcomes also depend on tool availability, retry policy, execution budget and task environment. Claims of general superiority require more evidence than favorable scores selected from one domain. A broad evaluation table is useful because it exposes weaknesses as well as strengths.

··········

ATRIA DAWN EXTENDS THE COMPETITION FROM LANGUAGE GENERATION TO VERIFIABLE TASK EXECUTION.

Shanghai AI Lab combines a 744-billion-parameter GLM-5.2 foundation with specialized evaluation for extended agentic workflows. The full model improves on the September Preview in JobBench, WideSearch, PaperBench, SWE-bench Pro and Terminal-Bench 2.1, though competitors remain stronger in several engineering and productivity tasks.

Two qualifications remain essential: the full model is documented and accessible through hosted services, but downloadable full-version weights remain forthcoming; and the official repository specifies 256K context despite some references to 512K.

Atria Dawn's significance lies in connecting frontier-scale language modeling with verifiable task execution. Its practical value will depend on reproducible completion, infrastructure cost, operational reliability and the eventual release of the full model weights.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org