Atria Dawn Preview review, pricing and limits

Atria Dawn Preview scores 96.0% on the vendor's DeepSearchQA research benchmark and runs locally via vLLM 0.23 or newer, ready since its September 2026 launch. It suits engineering and research teams that want an open, self-hostable agent model, not shops needing vision or audio input.

An agentic foundation model from Shanghai AI Lab, Atria Dawn Preview scores 59.6% on SWE-bench Pro since its September 2026 release. Built through post-training on Zhipu AI's GLM-5.2 base rather than trained from scratch, it targets long-horizon research and engineering work over short chat replies.

Provider: Shanghai AI Lab · Family: Atria Dawn

More about Shanghai AI Lab on HokAI

Context window: 256,000 tokens

Input modalities: text, tool-calls · Output: text, tool-calls

About Atria Dawn Preview

Atria Dawn Preview is a 744 billion parameter mixture of experts model Shanghai AI Lab published on Hugging Face on September 11, 2026, with an FP8-quantized checkpoint following on September 12 and a formal announcement on September 15. It is not trained from scratch: the model is built through agentic post-training on top of Zhipu AI's GLM-5.2 base, the same foundation model Z.ai shipped in June 2026. The release came from ATRIA, a cross-institutional research initiative Shanghai AI Lab runs with Fudan University's Natural Language Processing Lab, led by Fudan associate professor Tao Gui, with Shanghai AI Lab director Bowen Zhou among the roughly 140 listed co-authors on the accompanying arXiv paper, Atria Dawn: The Dawn of Agentic Superintelligence. It sits at the start of what the paper frames as a new Atria Dawn model family, and the lab presents it as a model built for agents from the post-training up, separate from its existing InternLM chat line.

Shanghai AI Lab's GitHub model card, re-read on October 9, 2026, publishes 16 benchmark results for Atria Dawn Preview next to six rival models, every figure run and reported by the lab itself. Atria Dawn Preview holds the highest score in that table on DeepSearchQA (96.0%), BrowseComp (92.5%), BFCL v4 (77.0%), AutomationBench (53.8%) and CyberGym (86.5%); those are the five first places the arXiv abstract claims without listing them. On every other row at least one rival scores higher, and the widest gaps are in coding and job-style delivery work.

Row by row, with the leading rival in brackets: for Discovery, WideSearch 81.9 (GPT-5.6 Sol, 83.3) and DeepResearch Bench II 51.1 (Claude Opus 5, 54.1). For Creation, MLE-bench Lite 86.2 (GPT-5.6 Sol, 88.9), SWE-bench Pro 59.6 (Claude Opus 5, 74.7) and Terminal-Bench 2.1 78.3 (Claude Opus 5, 90.2). For tool use, SkillsBench 66.4 (Qwen 3.8 Max, 66.7) and τ³-Bench Banking 41.2 (Qwen 3.8 Max, 55.2). For Delivery, Workspace-Bench 65.0, Workspace-Bench-Lite 68.2, GDPval 1583 Elo and JobBench 50.3, with Claude Opus 5 leading all four at 65.8, 70.1, 1768 and 68.0.

The six rivals in the lab's table are DeepSeek V4 Pro 0813, Kimi K3, Qwen 3.8 Max, GLM-5.3, GPT-5.6 Sol and Claude Opus 5, and the lab left a dash wherever it had no rival score. Against the other Chinese models the coding gap narrows: on SWE-bench Pro it edges DeepSeek V4 Pro 0813 (58.3%) and trails GLM-5.3 (60.3%), Kimi K3 (61.6%) and Qwen 3.8 Max (65.1%). Treat all of it as vendor-reported: aiweekly.co's launch review notes that no independent evaluator, Artificial Analysis included, had verified the model, and none had published a re-run by October 9, 2026.

Atria Dawn Preview ships with a 256,000 token context window, enough to hold a research paper, a codebase excerpt, or a long multi-step agent transcript in a single call. No independent long-context recall evaluation, such as a needle-in-haystack test, has been published for the model, so how well it holds up near the top of that window is unverified as of this writing.

The model card lists text-only input and output: there is no vision, audio, or video support documented, which sets it apart from the multimodal flagships most Western labs now ship by default. Its strength instead is tool use and agentic loop execution, reflected in the BFCL v4 function-calling benchmark and the AutomationBench score, and the vendor organizes its intended use around four areas it calls Discovery (research and evidence retrieval), Creation (software and application development), Delivery (document and data transformation) and Cybersecurity (vulnerability analysis and validation). Two agent harnesses, Codex and Claude Code, are listed as compatible integrations, meaning developers can point an existing agentic coding tool at Atria Dawn Preview instead of a proprietary model.

There is no published per-token pricing. The model ships as open weights under the MIT license, so the direct cost is whatever GPU infrastructure you provide to self-host it, not a metered API bill. Two hosted API endpoints went live alongside the release, an international one at api.atria-asi.ai and a China endpoint at discovery.intern-ai.org.cn, and the international endpoint's documentation, read on October 9, 2026, accepts Chat Completions, Anthropic-style Messages and Responses requests under a single model ID, Atria-Dawn-Preview. Neither endpoint listed a per-token price when re-checked that day.

Self-hosting requires SGLang version 0.5.13.post1 or newer, or vLLM version 0.23.0 or newer; both are recent releases, so older inference stacks will not load the checkpoint. Weights are published in BF16 and F32 for the base model, plus a separate FP8-quantized checkpoint aimed at lower-VRAM deployments, on both Hugging Face and ModelScope. aiweekly.co puts the download at 756GB to 1.5TB depending on precision, which makes self-hosting a multi-GPU job even at FP8. No managed cloud listing on AWS Bedrock, Google Vertex AI, or Azure was found.

No system card, red-teaming partner list, or refusal-rate benchmark specific to Atria Dawn Preview has surfaced yet. That is a gap relative to the lab's own public safety work elsewhere: its Intern-SafeWork stack and its published value-alignment and adversarial-robustness benchmarks cover the general InternLM line, and director Bowen Zhou chairs the AI safety working group of China's national cybersecurity standards committee, TC260, but none of that has yet been tied specifically to this model's release documentation.

Teams running long, multi-step agent loops such as literature review into experiment design, coding, or vulnerability analysis get a free, self-hostable model built specifically for that workflow rather than adapted from a chat model. Teams that need multimodal input, a managed API with a published SLA and price list, or benchmark scores independently confirmed beyond the vendor's own paper should look at a competitor such as DeepSeek V4 or Qwen 3.8 Max, or wait for third-party evaluation of Atria Dawn Preview to catch up.

The paper describes training through what it calls a Verifiable Experience Pipeline, connecting tool-mediated agent interactions to executable environments rather than relying only on static text corpora, and reports that human raters judged roughly one-third of completed AI-assisted tasks as infeasible without the model's help. Beyond that description, Shanghai AI Lab has not disclosed a training data cutoff date, an API input retention policy, or a stance on training future models on user inputs, and no SOC 2, HIPAA, or GDPR compliance statement was found for either hosted endpoint.

Pricing

Open-weight release under the MIT license, so the model itself costs nothing to download. No per-token price is published for either hosted endpoint, api.atria-asi.ai (international) or discovery.intern-ai.org.cn (China), as of October 9, 2026. Self-hosting means paying for your own GPUs instead; for a metered open-weight alternative with published rates, compare [DeepSeek V4 Pro 0813](/hub/models/deepseek-v4-pro-0813).

Key Features

• Agentic Post-Training on a Frontier Base: Adds agentic post-training on top of a large third-party mixture-of-experts foundation model rather than training from scratch.
• Long-Context Window for Agent Runs: Wide enough to hold a full research paper, codebase excerpt, or multi-step agent transcript in a single call.
• Four Agentic Focus Areas: Purpose-built workflows for Discovery (research), Creation (coding), Delivery (document and report generation) and Cybersecurity (vulnerability analysis).
• MIT-Licensed Open Weights: Both the base and an FP8-quantized checkpoint are downloadable from Hugging Face and ModelScope for commercial self-hosting.
• Dual-Region Hosted API: A separate international endpoint and a China endpoint both went live at launch, though neither publishes per-token pricing yet.

Pros

• Ships as fully open, MIT-licensed weights, so teams can self-host, fine-tune, or run air-gapped without a vendor contract.
• A wide context window and agentic post-training suit long research and coding sessions better than a short-context chat model.
• Free to run at whatever infrastructure cost you provide, with no per-token vendor markup on top.

Cons

• Trails on agentic coding: its SWE-bench Pro result is 15.1 points behind Claude Opus 5 in the lab's own table, and no outside evaluator has re-run any score yet.
• Text-only: no vision, audio, or video input, unlike multimodal frontier models from OpenAI, Google or Anthropic.
• No managed hosted pricing, SLA, or enterprise support published as of the September 2026 launch, so production reliability depends on self-hosting.

Benchmarks

• SWE-bench Pro: 59.6% vendor-reported · 09 Oct 2026 — Harder, longer real-repository coding tasks, % solved.
• Terminal-Bench 2.1: 78.3% vendor-reported · 09 Oct 2026 — Multi-step tasks completed in a real command line, % solved.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Atria Dawn Preview Official Page