753B checkpoint / ~40B active (MoE; post-trained from the 744B-A40B GLM-5.2 foundation)
256K supported release context (checkpoint max_position_embeddings: 1,048,576)
MIT
Open-weight preview agentic model developed by the Shanghai Artificial Intelligence Laboratory for long-horizon scientific research, software engineering, digital work, and tool-mediated problem solving. Atria Dawn Preview is a post-trained derivative of Z.AI's GLM-5.2 foundation rather than a newly pretrained base model. Its shipped configuration retains the 78-layer GLM MoE DSA architecture, comprising three initial dense layers followed by 75 sparse layers, a 6,144-dimensional hidden state, 256 routed experts plus one shared expert, eight routed experts selected per token, 64 attention heads, compressed query and key/value representations, a 2,048-token sparse-attention index, shared indexers across groups of sparse layers, and one Multi-Token Prediction layer. Hugging Face calculates approximately 753B checkpoint parameters, while the developers describe the underlying GLM-5.2 foundation as 744B parameters; the inherited architecture activates approximately 40B parameters per token. The model was trained through a Verifiable Experience Pipeline in which agents interact with real execution environments, call tools, produce intermediate artifacts, respond to environmental feedback, and have final outcomes checked through executable tests, experiment metrics, file and application state, geometric checks, source evidence, or human-defined criteria. Trajectory curation removes incomplete, contradictory, duplicate, and behaviorally invalid examples, while failure analysis informs additional task construction and environment refinement. The developers organize its target capabilities into discovery, creation, delivery, and authorized cybersecurity work. They report leading scores within their comparison table on AutomationBench at 53.8, BFCL v4 at 77.0, CyberGym at 86.5, DeepSearchQA at 96.0, and BrowseComp at 92.5. Additional reported results include 66.4 on SkillsBench, 65.0 on Workspace-Bench, 86.2 HumanRank on MLE-bench Lite, 78.3 on Terminal-Bench 2.1, 59.6 on SWE-bench Pro, 51.1 on DeepResearch Bench II, and 1583 Elo on GDPval. The accompanying report analyzes 769 development-task records from 56 participants; among recorded methods and decisions, agents proposed 64.6 percent while humans retained the final choice in 85.5 percent. The public release and evaluation harnesses use a 256K-token context, although the checkpoint configuration declares 1,048,576 maximum positions. Official BF16 and FP8 checkpoints occupy approximately 1.51 TB and 756 GB respectively, with deployment support documented for SGLang 0.5.13.post1 or newer and vLLM 0.23.0 or newer. The model accepts text input and produces text output.