Shanghai AI Laboratory has posted the weights of Atria Dawn Preview, its new-generation agentic model, on Hugging Face and ModelScope — MIT license, built on a 744B-parameter MoE GLM-5.2 foundation, 256K context, with an FP8-quantized variant released alongside. The model skips the chatbot persona entirely; its one pitch is pushing open-ended research tasks toward results that are executable, verifiable, and reproducible.
Verification built into the training loop
At Atria's core is a Verifiable Experience Pipeline: every training task connects to a real execution environment, where the model observes state, calls tools, produces intermediate artifacts, and adapts to feedback. Final outcomes are verified through external signals — passing tests, metrics, file state, geometric structure, or source evidence. Only experience that ties together the task, its trajectory, its artifacts, and verification evidence gets absorbed as reusable capability. The team summarizes this as a three-way split: the model internalizes past experience, the harness organizes the process at hand, and the environment determines the consequences of each action. In effect, it replaces the old "bolt agent skills on with prompt engineering" route with internalization during training.
Benchmarks: agent tasks lead, engineering delivery lags
The official comparison table pulls in six rivals: DeepSeek V4 Pro 0813, Kimi K3, Qwen 3.8 Max, GLM 5.3, GPT 5.6 sol, and Claude Opus 5. Across 16 benchmarks, Atria posts the highest score in its row on five: AutomationBench 53.8 (next best 49.7), BFCL v4 77.0, CyberGym 86.5, DeepSearchQA 96.0, and BrowseComp 92.5. Long-horizon agent work — browser-style deep search and cybersecurity — is where the lead is widest.
The same table also exposes the weaknesses: SWE-bench Pro 59.6 against Opus 5's 74.7; JobBench 50.3 versus 68.0; Terminal-Bench 2.1 78.3 versus 90.2. Strong on research-exploration tasks, weak on engineering-delivery tasks — a clear profile that the official page does not hide, which counts for more credibility than a press-release launch.
One demo: from 100GB of weather data to a global forecast system
The hardest demo on the site: in an environment with web search disabled, given more than 100GB of global weather data, the model designed a ViT network with over 0.4 billion parameters, trained it for 45,000 steps, modeled the evolution of 69 weather variables, and produced a system that forecasts a week of global weather within a minute — outperforming NVIDIA's classic FourCastNet on some prediction metrics (per the official claim). A separate demo builds a MiniOS from scratch in 20 minutes.
So what
It's a preview, text-input only, and the comparison table is self-reported — take the discounts. Two things are worth watching. First, 744B MoE-class weights fully open under MIT, with deployment docs covering SGLang, vLLM, Codex, and Claude Code out of the box. Second, as the agent race shifts from chat scores to task-delivery scores, where the verification signal comes from decides how much a score is worth — Atria putting verification inside the training loop matters more than any single leaderboard number.
Refs: Atria site · HF model card