Atria Dawn
modelYour notes
Shanghai AI Laboratory's Atria line: a "foundation agentic language model" for research and engineering workflows, built by post-training Z.ai's 744B-parameter GLM-5.2 MoE. The card says so directly ("Built on the 744B-parameter MoE GLM-5.2 foundation model"), and the config.json is GLM-5.2's (GlmMoeDsaForCausalLM, 78 layers, hidden size 6,144, 256 routed experts plus one shared, 8 active per token), so the parameter count belongs to Z.ai's pretraining and is not recorded as Shanghai AI Lab's. The Preview shipped on 11 September 2026 with a 256K context, an FP8 sibling and, on 15 September, an Ascend W8A8 build. The full release followed on 23 October 2026 (internlm/Atria-Dawn plus an FP8 build, 353 BF16 safetensors shards, about 1.5 TB). It doubles the stated context to 512K, is text-input only, and keeps the MIT license. The card frames the model around four kinds of agentic work: discovery, creation, delivery, and cybersecurity.
The accompanying 143-author paper covers the Preview only. It describes a Verifiable Experience Pipeline that ties tool-mediated interactions to executable environments and externally verified outcomes, reports the Preview as competitive with frontier agents across 16 benchmarks with the highest reported score on five, and treats the model's own development as a case study in human–AI collaboration, analysing 769 task records from the R&D process. No paper or changelog explains what changed between Preview and full release.
The release card's table, all self-reported with the lab's own harnesses, puts the full release slightly ahead of the Preview on most rows: AutomationBench 54.7 (Preview 53.8), CyberGym 86.8 (86.5), SWE-bench Pro 61.7 (59.6), Terminal-Bench 2.1 81.0 (78.3), BFCL v4 78.4 (77.0), DeepSearchQA 96.5 (96.0), MLE-bench Lite 88.2 (86.2) and JobBench 57.5 (50.3). BrowseComp slipped from 92.5 to 91.3 and NL2Repo-Bench from 51.4 to 50.0. It leads the compared column (DeepSeek V4 Pro 0813, Kimi K3, Qwen3.8 Max, GLM-5.3, GPT-5.6 Sol, Claude Opus 5) on the search, tool-use and automation rows, but still trails the coding leaders by a wide margin: SWE-bench Pro 61.7 against Opus 5's 76.5, Terminal-Bench 2.1 81.0 against 90.2, and on the newly added DeepSWE row 49.3 against 72.6, below every compared model. Neither checkpoint had been scored by Artificial Analysis at filing (no AA model page exists for either).
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| AutomationBench | 54.7 | — |
| CyberGym | 86.8 | — |
| DeepSearchQA | 96.5 | — |
| BrowseComp | 91.3 | — |
| BFCL v4 | 78.4 | — |
| SWE-bench Pro | 61.7 | — |
| Terminal-Bench 2.1 | 81.0 | — |
| DeepSWE | 49.3 | — |
| MLE-bench Lite | 88.2 | — |
| GDPval | 1594.2 | — |
Variants
| Name | Parameters | Notes |
|---|---|---|
| Atria Dawn | — | full release, 2026-10-23; 512K context, text-only input; BF16 and FP8; post-train of GLM-5.2 |
| Atria Dawn Preview | — | 2026-09-11; 256K context; BF16, FP8 and (2026-09-15) Ascend W8A8; DeepSearchQA 96.0, BrowseComp 92.5, BFCL v4 77.0, AutomationBench 53.8, SWE-bench Pro 59.6, Terminal-Bench 2.1 78.3 |