Atria Dawn Preview: Shanghai AI Lab’s 744B open agent model
Confirmed
The takeaway
Its own table puts it first on BrowseComp (92.5), DeepSearchQA, BFCL v4, AutomationBench and CyberGym, but behind Claude Opus 5 on SWE-bench Pro and JobBench.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 17 October 2026
Your AI and this story
- GPT-6 Astra134 days after its cutoff
- Claude Opus 5.573 days after its cutoff
- Gemini 4 Argon164 days after its cutoff
- Grok 4.7103 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 73 days before it.
Key facts
- Hugging Face repo internlm/Atria-Dawn-Preview created Sept 11, 2026; MIT license; ~753B parameters in safetensors; text-only input; 256K context; FP8 build Sept 12 and an Ascend w8a8 build Sept 15; also on ModelScope
- Base: 744B-parameter MoE GLM-5.2 (Z.ai); model card cites arXiv 2609.15818
- Vendor benchmarks (model card): BrowseComp 92.5 (GPT-5.6 Sol 92.2, Claude Opus 5 90.8); DeepSearchQA 96.0; BFCL v4 77.0; AutomationBench 53.8; CyberGym 86.5; SWE-bench Pro 59.6 (Claude Opus 5 74.7); Terminal-Bench 2.1 78.3 (Opus 5 90.2); GDPval 1583 (Opus 5 1768); JobBench 50.3 (Opus 5 68.0)
- Access: API at api.atria-asi.ai (international) and Intern Discovery token plan (China); local serving with SGLang ≥0.5.13.post1 or vLLM ≥0.23.0; the card documents Codex CLI setup via the Responses API
- Full version (Pandaily, Oct 17, 2026; not confirmed on the lab’s pages yet): 512K context, higher JobBench, WideSearch and PaperBench scores, weights ‘coming soon’
What happened
Shanghai Artificial Intelligence Laboratory, the state-backed lab behind the InternLM and Intern-S models, published Atria Dawn Preview with little fanfare on Sept 11, 2026. It is not trained from scratch: it is post-trained on Z.ai’s open 744B MoE GLM-5.2. The lab pitches it as an agent that takes “research questions to verifiable results”, covering discovery (deep research), creation (software, ML systems), delivery (reports, slides) and cybersecurity in authorised environments.
On the lab’s own table it leads the search and tool-use benchmarks it reports, roughly ties GPT-5.6 Sol on BrowseComp, and trails Claude Opus 5 by wide margins on SWE-bench Pro, Terminal-Bench 2.1, GDPval and JobBench. These are vendor numbers.
Why it matters
It shows Chinese labs building frontier agent models on each other’s open weights: a government lab fine-tuning a startup’s 744B base and releasing it under MIT, with its own API and documented use inside OpenAI’s Codex CLI.
Full version (Oct 17)
Pandaily reported on Oct 17 that a full version had been released with a 512K context and better JobBench, WideSearch and PaperBench scores, with weights still marked “coming soon”. We could read only the headline and a search snippet, and the lab’s Hugging Face page listed no new repository that evening, so treat these details as unconfirmed.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 2 primary, 2 press
Primary
- Hugging Face: internlm/Atria-Dawn-Preview (model card)huggingface.co, official
- Atria API docsapi.atria-asi.ai, docs
Press
Changes
- Filed (found while checking the Oct 17 Pandaily report; the September preview had not been covered)