模型發布
Atria Dawn Preview Releases 744B MoE Weights, Trained for Long-Horizon Research Agents in Verifiable Environments
The Shanghai AI Laboratory has released Atria Dawn Preview, along with FP8 weights and inference recipes, positioning it as a full-loop agent capable of searching, coding, and running experiments. The organization claims the best reported results on five of 16 benchmarks, though model scale, test-time compute, and its in-house evaluation setup still limit comparability.
On September 14, the Shanghai AI Laboratory released Atria Dawn Preview, a text-based agent model built on the 744B-parameter GLM-5.2 MoE base model. The team also released full-precision and FP8 weights under the MIT License, with a 256K-token context window and deployment paths for SGLang 0.5.13.post1 and vLLM 0.23.0 or later. This makes it more than a hosted API: teams with multi-node GPU resources can download the checkpoint and retain control over tools, sandboxes, and data boundaries.
The training centers on a “Verifiable Experience Pipeline.” Rather than learning solely from textual answers, the model searches for information, writes code, runs experiments, analyzes results, and recovers from failures in executable environments, with feedback derived from externally verifiable outcomes. The paper groups its capabilities into discovery, creation, delivery, and cybersecurity, aiming to equip the model for long-horizon, open-ended tasks rather than only single-turn question answering or isolated coding problems.
Across 16 official evaluations covering search, coding, tool use, office productivity, and cybersecurity, the team claims that Atria achieved the highest reported score on five benchmarks. These include 92.5 on BrowseComp, 53.8 on AutomationBench, and 86.5 on CyberGym. However, it still trails some closed-source models on tests including SWE-bench Pro, Terminal-Bench 2.1, and GDPval. The paper also analyzes 769 R&D task records from 56 participants. Participants reported that roughly one-third of the AI-assisted tasks they completed would have been difficult to finish without AI, but these were internal team work records and cannot be directly generalized to developer productivity overall.
Engineering teams should watch two issues next. First, even though a 744B MoE model activates only a subset of experts at a time, weight storage requirements and cross-GPU communication still make local deployment highly demanding. Second, the current benchmark tables were produced primarily by the releasing team, and independent reruns using the same agent framework, tool budget, and test-time compute are still lacking. The Preview designation also suggests that the API, prompt format, and inference support may continue to change.