Models
Shanghai AI Laboratory releases open-weight agentic model Atria Dawn Preview
Shanghai Artificial Intelligence Laboratory has released Atria Dawn Preview, an open-weight agentic model built on the GLM-5.2 mixture-of-experts base, under an MIT license. A technical report posted to arXiv on Sept. 14 describes its training approach and reports results on 16 benchmarks.
The model card on Hugging Face, where the weights sit under the lab's InternLM organization, says Atria Dawn Preview builds on GLM-5.2, which it describes as a 744-billion-parameter mixture-of-experts model; the listing shows about 753 billion parameters. The model accepts text only, supports a 256,000-token context window and comes with an FP8 variant. The card points to SGLang and vLLM for local deployment and also lists hosted API access. The GitHub repository says both code and weights are MIT-licensed.
The report, credited to more than 140 authors, says training used what it calls a Verifiable Experience Pipeline, which ties tool use to executable environments with externally checked outcomes. The authors say the model posted the highest reported score on five of the 16 benchmarks; the card lists 96.0 on DeepSearchQA and 59.6 on SWE-bench Pro. These are the lab's own figures. Startup Fortune, which reported that the repository appeared on Sept. 11 without an announcement, noted that no outside group has reproduced them.
The paper also examines 769 task records from 56 participants, who rated about a third of completed AI-assisted tasks as infeasible without AI.
Source details
- Source
- arXiv
Source reporting
Read the original reporting and research behind this briefing.
