How good is Atria Dawn Preview compared with DeepSeek V4 Pro and GLM-5.3?

On the lab's own table it leads on agent tool-use and discovery: BFCL v4 77.0 (DeepSeek V4 Pro 71.4, GLM-5.3 74.1), AutomationBench 53.8 (V4 Pro 41.7, GLM-5.3 49.2), DeepSearchQA 96.0 and CyberGym 86.5. It trails on coding: SWE-bench Pro 59.6 versus 65.1 for Qwen 3.8 Max and 74.7 for Claude Opus 5, and Terminal-Bench 2.1 78.3 versus 89.3 for Qwen 3.8 Max and 90.2 for Opus 5. These are vendor-reported, single-run numbers.

Is Atria Dawn Preview free to use?

Yes, in preview. The weights are MIT-licensed on Hugging Face and ModelScope, and an OpenAI-compatible hosted API at api.atria-asi.ai launched as a free preview (a separate China endpoint runs on intern-ai.org.cn). Pricing for the hosted service after preview has not been announced as of September 16, 2026. Self-hosting a 744B MoE needs a multi-GPU node: the lab documents SGLang v0.5.13.post1+ and vLLM v0.23.0+ using the GLM-5.2 recipes.

Can I use Atria Dawn Preview in Codex or other coding agents?

Yes. The model card documents a Codex setup: add a custom provider in ~/.codex/config.toml pointing at https://api.atria-asi.ai/v1 with wire_api = 'responses', and a model catalog file that declares input_modalities as text only and a 256,000-token context window, because Codex otherwise assumes multimodal input and the endpoint rejects images. Any OpenAI-compatible harness (OpenHands, OpenClaw, Cline) can point at the same endpoint.

Quick Answer

What Is Atria Dawn Preview? Shanghai AI Lab's 744B Agent

Q: What is Atria Dawn Preview?

Atria Dawn Preview is an open-weight agentic model from the Shanghai Artificial Intelligence Laboratory (the InternLM team), released on Hugging Face on September 11, 2026 with an arXiv technical report (2609.15818, 143 authors) on September 14. It is post-trained on Z.ai's 744B-parameter mixture-of-experts GLM-5.2 base, ships under the MIT license in BF16 and FP8 variants, has a 256K-token context window, accepts text only, and is built for long-horizon research and engineering tasks: deep research, software creation, document delivery and authorized security work.

Published: September 16, 2026

The short answer

Atria Dawn Preview is a 744-billion-parameter, MIT-licensed agentic model from the Shanghai Artificial Intelligence Laboratory, released quietly on September 11, 2026 and documented in a 143-author arXiv report three days later. It is not a chat model. It is post-trained on top of Z.ai’s GLM-5.2 mixture-of-experts base to run long loops of research, tool use, code, experiment and recovery, and the lab positions it for “scientific automation and office work.” A free hosted API and the open weights are both live as of September 16, 2026.

Key facts

	Atria Dawn Preview
Developer	Shanghai Artificial Intelligence Laboratory (InternLM team), brand “Atria ASI”
Released	Weights on Hugging Face September 11, 2026; arXiv 2609.15818 September 14, 2026
Base model	GLM-5.2 (Z.ai), 744B-parameter MoE
License	MIT
Variants	Atria-Dawn-Preview (BF16), Atria-Dawn-Preview-FP8
Context	256K tokens
Modality	Text in, text out; no image input
Hosted API	api.atria-asi.ai (international, OpenAI-compatible incl. Responses API), intern-ai.org.cn (China); free preview
Local serving	SGLang ≥ v0.5.13.post1, vLLM ≥ v0.23.0 (GLM-5.2 recipes)

What it is designed to do

The model card describes four target dimensions, and the benchmark table is organized around them:

Discovery — retrieving and organizing evidence, deep research, turning a research question into an executable plan. Scores: DeepSearchQA 96.0, BrowseComp 92.5, WideSearch 81.9, DeepResearch Bench II 51.1.
Creation — building software, apps, games, visualizations and ML systems. MLE-bench Lite 86.2, SWE-bench Pro 59.6, Terminal-Bench 2.1 78.3.
Delivery — turning documents, data and requirements into reports and presentations. Workspace-Bench 65.0, GDPval 1583, JobBench 50.3.
Cybersecurity — analyzing, validating, fixing and re-validating vulnerabilities in authorized environments. CyberGym 86.5, the top score in the table.

Tool-use rows are where it stands out most: BFCL v4 77.0 and AutomationBench 53.8 beat every listed comparator, including GPT-5.6 Sol and Claude Opus 5 where those have numbers.

Where it is weaker

Coding is the clear gap. Claude Opus 5 posts 74.7 on SWE-bench Pro and 90.2 on Terminal-Bench 2.1 against Atria’s 59.6 and 78.3; Qwen 3.8 Max scores 65.1 and 89.3. Atria also trails on GDPval (1583 vs Opus 5’s 1768) and JobBench (50.3 vs 68.0). The lab is explicit that this is a preview and that the benchmark figures are its own runs; there is no independent Artificial Analysis or third-party harness score yet. The full comparison is in Atria Dawn Preview vs DeepSeek V4 Pro vs GLM-5.3 vs Kimi K3.

The human study in the paper

Unusually, the technical report includes 769 task records from 56 people who used the model during development. Participants rated roughly one-third of the AI-assisted tasks they completed as infeasible without the model. The authors frame this as a participant judgment rather than evidence of autonomy and stress that humans kept final decision authority on every task, which matches the model’s “verifiable results” framing: the loop ends in something a person can check.

How to run it

Hosted (fastest): get a key from the api.atria-asi.ai console and call it as an OpenAI-compatible endpoint. The service supports both Chat Completions and the Responses API.

In Codex: the model card ships a working config. In ~/.codex/config.toml:

model = "Atria-Dawn-Preview"
model_provider = "atria"
[model_providers.atria]
name = "Atria"
base_url = "https://api.atria-asi.ai/v1"
env_key = "ATRIA_API_KEY"
wire_api = "responses"

Then add a model catalog JSON declaring "input_modalities": ["text"] and "context_window": 256000; without it Codex attaches images by default and the endpoint returns 400 Atria-Dawn-Preview is not a multimodal model.

Self-hosted: the FP8 checkpoint is the practical one. Expect the same footprint as GLM-5.2/5.3 FP8, which the lab and Z.ai document on multi-GPU nodes such as an 8×B200/B300 server or a DGX Station cluster; SGLang v0.5.13.post1+ or vLLM v0.23.0+ with the published GLM-5.2 recipes.

Why it matters

Second major open agentic model on the GLM-5.2 base. Z.ai’s own GLM-5.3 and now Atria both build on it, making GLM-5.2 the de facto open foundation for 700B-class agent post-training in 2026.
MIT license with no open-weights delay. Muse Spark 1.3 and GLM-5.3 Flash have promised open weights that have not shipped; Atria shipped weights first and the paper second.
Free preview API. For teams evaluating open agent models against DeepSeek V4 Pro (off-peak $0.66/$1.98 per MTok) or Kimi K3 ($3/$15), the marginal cost of trying Atria today is zero.
State-lab provenance. Shanghai AI Laboratory is a government-backed institute; enterprises with data-residency or procurement rules should treat the hosted endpoint like any other China-hosted API and self-host if that matters.

What to watch

Independent scores on SWE-bench Pro, Terminal-Bench 2.1 and the Artificial Analysis Intelligence Index.
Post-preview pricing for api.atria-asi.ai.
Whether a “Dawn” GA release adds image input; the text-only limit blocks most computer-use work.