Language modelPreview
Atria Dawn Preview
Shanghai Artificial Intelligence Laboratory
- Released
- -
- Data date
- October 7, 2026
- Input
- Text
- Output
- Text
Atria Dawn Preview is a text model from Shanghai Artificial Intelligence Laboratory for research and engineering agents. Its model card says it builds on the 744B MoE GLM-5.2 foundation; that figure describes the base model, not a published parameter count for Atria Dawn.
Specifications and access
| Specification | Value and source |
|---|---|
| Provider model ID | Atria-Dawn-PreviewSource |
| Access | API and local deploymentSource |
| Input modalities | TextSource |
| Output modalities | TextSource |
| License | MITSource |
| Reported context | Model variants are listed with 256K context; an exact token count is not stated.Source |
| Foundation model | Built on the 744B MoE GLM-5.2 foundation model; this is not a parameter count for Atria Dawn.Source |
Page 1 of 2
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Terminal-Bench 2.1 (Show measurement, test conditions, and source)
- Source value
- 78.3
- Score
- 78.3
- Metric
- average pass rate
- Unit
- %
- Benchmark version
- 2.1
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Claude Code v2.1.224 unless otherwise specified, 12x timeout capped at 12 hours, up to 500 turns, mean pass rate over four runs; GPT models use Terminus 2.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
JobBench (Show measurement, test conditions, and source)
- Source value
- 50.3
- Score
- 50.3
- Metric
- mean task percentage score
- Unit
- %
- Category
- professional-work
- Direction
- Higher is better
- Sample
- 65
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official JobBench OpenCode harness, main split of 65 tasks, 7,200-second timeout and two attempts maximum, 10 concurrent workers, Grok-4.3 judge; arithmetic mean percentage, API failures score zero.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
AutomationBench (Show measurement, test conditions, and source)
- Source value
- 53.8
- Score
- 53.8
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Benchmark version
- 1.0.6
- Category
- agentic
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AutomationBench v1.0.6, with null-type handling fix from PR #13.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
BFCL v4 (Show measurement, test conditions, and source)
- Source value
- 77
- Score
- 77
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Benchmark version
- harness commit 6ea5797
- Category
- tool-use
- Direction
- Higher is better
- Sample
- 5217
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Official BFCL v4 harness commit 6ea5797, 5,217 cases, temperature 0.001, 120-second request timeout, at most two attempts, Serper for web search, and official 10/10/10/30/40 weighting.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
CyberGym (Show measurement, test conditions, and source)
- Source value
- 86.5
- Score
- 86.5
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Category
- cybersecurity
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Claude Code v2.1.119, 2 CPUs and 4 GB per task, 4 hours 10 minutes timeout, thinking effort high.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
DeepSearchQA (Show measurement, test conditions, and source)
- Source value
- 96
- Score
- 96
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Category
- research
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. In-house search, visit and Python tools, up to 500 steps per task, 256K context, maximum 64K output tokens per generation.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
Workspace-Bench-Lite (Show measurement, test conditions, and source)
- Source value
- 68.2
- Score
- 68.2
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Category
- workplace
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AgentCompass English Lite dataset, 2 CPUs and 8 GiB per task, two-hour limit; GPT models use Codex and other models Claude Code; Doubao Seed 2.0 Lite judge.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
BrowseComp (Show measurement, test conditions, and source)
- Source value
- 92.5
- Score
- 92.5
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Category
- research
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. In-house search, visit and Python tools, up to 500 steps per task, 256K context, maximum 64K output tokens; discard-all context management.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
SkillsBench (Show measurement, test conditions, and source)
- Source value
- 66.4
- Score
- 66.4
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Category
- tool-use
- Direction
- Higher is better
- Sample
- 79
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. OpenHands within AgentCompass on a self-contained 79-task subset excluding multimodal tasks; mean of three runs.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
Workspace-Bench (Show measurement, test conditions, and source)
- Source value
- 65
- Score
- 65
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Category
- workplace
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. AgentCompass English Full dataset, 2 CPUs and 8 GiB per task, two-hour limit; GPT models use Codex and other models Claude Code; Doubao Seed 2.0 Lite judge.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
MLE-bench Lite (Show measurement, test conditions, and source)
- Source value
- 86.2
- Score
- 86.2
- Metric
- HumanRank score
- Unit
- No unit provided
- Category
- machine-learning
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Claude Code, 12-hour time limit per evaluation; reported as HumanRank score.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
WideSearch (Show measurement, test conditions, and source)
- Source value
- 81.9
- Score
- 81.9
- Metric
- reported score; unit not stated by paper
- Unit
- No unit provided
- Category
- research
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Atria Dawn technical report
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. In-house search, visit and Python tools, up to 500 steps per task, 256K context, maximum 64K output tokens per generation.
- Context
- Vendor-reported paper results for the exact Atria Dawn Preview profile. Scores across different benchmarks or harnesses are not a common cohort and should not be read as directly comparable measurements.
Page 1 of 2
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not found in the inspected sources | independent-reproduction Original wording No independent evaluator with numeric results for the exact Atria Dawn Preview checkpoint was verified in this pass; all included results are from the provider's paper. |
| Additional source | Atria Dawn Preview repository (retrieved October 7, 2026) · Atria Dawn Preview repository · Editorial description reviewed October 7, 2026 |
| Additional source | Atria Dawn technical report (retrieved October 7, 2026) |