中文 ｜ English

Atria Dawn Preview: From Research Questions to Verifiable Results

Introduction · Model Downloads · Evaluation · Deployment · License · Contact

---

🌅 Introduction

Atria Dawn Preview is a preview release of a new-generation agentic model developed by the Shanghai Artificial Intelligence Laboratory. Built on the 744B-parameter MoE GLM-5.2 foundation model, it targets research and engineering scenarios that demand continuous environmental understanding, tool use, and multi-step task completion.

The model is designed to drive open-ended problems toward executable, verifiable, and reproducible results. By combining task objectives with environmental feedback, it supports the full loop of problem analysis, solution design, tool use, code implementation, experiment execution, result analysis, and failure recovery.

Atria Dawn Preview empowers agentic tasks across four dimensions, with a particular focus on end-to-end delivery in real-world productivity scenarios such as scientific automation and office work:

📥 Model Downloads

📊 Evaluation Results

We conducted a comprehensive evaluation of Atria Dawn Preview across search, coding, tool use, productivity, and security benchmarks.

Bold marks the best score in each row. – indicates the result is unavailable.

🚀 Deployment & Online Access

Atria Dawn Preview supports both local deployment and hosted access.

Online Access

Use the service endpoint corresponding to your region.

Local Deployment

Codex

Add a custom provider to ~/.codex/config.toml. Codex uses the Responses API.

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"

Restricting input to text only

Atria-Dawn-Preview accepts text input only. By default Codex assumes every model is multimodal and will attach images from -i/--image or a TUI paste, which the endpoint rejects with 400 Atria-Dawn-Preview is not a multimodal model. Declare the model's modalities so Codex strips image input on the client side instead.

Step 1 — Create a model catalog file

Save this as ~/.codex/atria-catalog.json:

{
  "models": [
    {
      "slug": "Atria-Dawn-Preview",
      "display_name": "Atria-Dawn-Preview",
      "base_instructions": "You are a coding agent running in the Codex CLI. You collaborate with the user in a shared workspace to accomplish their software engineering goals.\n\nYou can only receive text input. Images, screenshots, PDFs, and other binary attachments are not available to you. If the user refers to an attachment you cannot see, say so plainly and ask them to paste the relevant text instead.",
      "supported_reasoning_levels": [
        { "effort": "low", "description": "Fast responses with lighter reasoning" },
        { "effort": "medium", "description": "Balances speed and reasoning depth" },
        { "effort": "high", "description": "Greater reasoning depth for complex problems" }
      ],
      "shell_type": "unified_exec",
      "visibility": "list",
      "supported_in_api": true,
      "priority": 1,
      "support_verbosity": false,
      "truncation_policy": { "mode": "tokens", "limit": 10000 },
      "experimental_supported_tools": [],
      "context_window": 256000,
      "max_context_window": 256000,
      "input_modalities": ["text"]
    }
  ]
}

"input_modalities": ["text"] is the setting that disables multimodal input.

Set context_window / max_context_window to the model's real limit — Codex uses these to budget the prompt and decide when to auto-compact. Without them it falls back to a conservative default, which wastes usable context.

All other fields are required by the parser — omitting any one fails with missing field <name> and Codex will not start.

Step 2 — Point your config at it

model = "Atria-Dawn-Preview"
model_provider = "atria"
model_catalog_json = "~/.codex/atria-catalog.json"
[features]
view_image = false
[model_providers.atria]
name = "Atria"
base_url = "https://api.atria-asi.ai/v1"
env_key = "ATRIA_API_KEY"
wire_api = "responses"

features.view_image = false is optional — it removes the image-viewing tool so the model doesn't attempt a call that would be refused.

Important: model_catalog_json replaces the model catalog, it does not merge with it. Any model not listed in your file falls back to default metadata — which assumes multimodal input — and logs warning: Model metadata for <slug> not found. If you switch models with -m or by editing model, add that model to the same file, or the text-only restriction will not apply to it.

Requires Codex CLI 0.154.0 or later.

Claude Code

#!/usr/bin/env python3
"""
PreToolUse hook: block the Read tool from reading PDF and image files.
"""
import json
import sys
IMAGE_EXTENSIONS = (
    ".apng",
    ".avif",
    ".bmp",
    ".gif",
    ".heic",
    ".heif",
    ".ico",
    ".jfif",
    ".jpeg",
    ".jpg",
    ".jxl",
    ".png",
    ".svg",
    ".tif",
    ".tiff",
    ".webp",
)
def main():
    hook_input = json.loads(sys.stdin.read())
    file_path = hook_input.get("tool_input", {}).get("file_path", "")
    if file_path.lower().endswith(".pdf"):
        reason = "Reading PDF files with the Read tool is not allowed."
    elif file_path.lower().endswith(IMAGE_EXTENSIONS):
        reason = "Reading image files with the Read tool is not allowed."
    else:
        sys.exit(0)
    output = {
        "hookSpecificOutput": {
            "hookEventName": "PreToolUse",
            "permissionDecision": "deny",
            "permissionDecisionReason": reason,
        }
    }
    print(json.dumps(output, ensure_ascii=False))
    sys.exit(0)
if __name__ == "__main__":
    main()

Please save the above code script as ${Target_dir}/block_pdf_image_read.py, and be sure to use an absolute path.

Add it to ~/claude_dir/settings.json. This can implement interception of multimodal inputs, such as images and PDFs, in PreToolUse.

"hooks": {
    "PreToolUse": [
      {
        "matcher": "Read",
        "hooks": [
          {
            "type": "command",
            "command": "python3 ${Target_dir}/block_pdf_image_read.py",
            "timeout": 5
          }
        ]
      },
      {
        "matcher": "",
        "hooks": []
      }
    ]
}

Kimi Code

Add a custom provider to ~/.kimi-code/config.toml. Kimi Code talks to Atria over the OpenAI-compatible Chat Completions API.

This example targets Kimi Code 0.43.1 and later. Check your version first; config fields can vary between versions.

kimi --version
cp -a ~/.kimi-code/config.toml ~/.kimi-code/config.toml.bak

1. Write the provider and model into your user config

Merge this into ~/.kimi-code/config.toml. Keep default_model at the top level, before the tables. Update existing entries instead of duplicating fields or tables. Your atr_... key goes on the api_key line — keys are created in the Atria console and displayed only once.

default_model = "atria/Atria-Dawn-Preview"

[providers.atria]
type = "openai"
# must include /v1 — Kimi appends /chat/completions verbatim
base_url = "https://api.atria-asi.ai/v1"
# literal value only: no ${VAR} expansion, and OPENAI_API_KEY is not read
api_key = "ATRIA_API_KEY"

[models."atria/Atria-Dawn-Preview"]
provider = "atria"
# case-sensitive wire name, separate from the section key above (a local alias)
model = "Atria-Dawn-Preview"
display_name = "Atria Dawn Preview"
# required — without it the entry silently fails to load. Keep at the real limit
# so context budgeting and compaction work.
max_context_size = 256000
# the max_tokens sent on the wire. Without it Kimi sends max_context_size
# instead, which Atria rejects (valid range is 1-65536).
max_output_size = 65536
# text-only model: adding image_in fails with "is not a multimodal model".
# Declare tool_use explicitly, or Kimi treats capabilities as unknown.
# "thinking" is what turns reasoning on — without it the effort is forced to
# "off" and no reasoning_effort is sent.
capabilities = ["tool_use", "thinking"]
# without support_efforts the effort stays the opaque "on", which sends no
# reasoning_effort at all. Listing them makes /think and the picker work.
support_efforts = ["low", "medium", "high", "xhigh", "max"]
default_effort = "max"
# required to allow turning thinking off. Omitting it fails with "reasons by
# default but declares no off effort".
off_effort = "none"

Keep the file private, since the key is stored in plaintext:

chmod 600 ~/.kimi-code/config.toml

2. Restart and verify

Validate the config, then send a text-only task.

kimi doctor
kimi provider list
kimi

Expected:

OK config.toml  /home/<user>/.kimi-code/config.toml
All checked config files are valid.

atria  type=openai  models=1  source=inline
Default model: atria/Atria-Dawn-Preview

If atria is missing from the provider list, the model entry failed to resolve — most often an absent max_context_size.

📄 License

The code and model weights in this repository are released under the MIT License.

📮 Contact

Questions, suggestions, or collaboration ideas are welcome:

• 🐛 Issues & feature requests — open an issue on GitHub
• 💬 Community — join us on Discord
• 🐦 Updates — follow @AtriaASI on X

Citation

@misc{guo2026atriadawndawnagentic,
      title={Atria Dawn: The Dawn of Agentic Superintelligence}, 
      author={Honglin Guo and Tao Gui and Yicheng Chen and Guanting Dong and Qiming Ge and Yuyang Hu and Zixian Huang and Jiajie Jin and Alexander Lam and Yining Li and Jiahang Lin and Yanjiang Liu and Xinyu Lu and Haijun Lv and Junlin Shang and Qisheng Su and Guoqiang Wang and Rui Wang and Zhecan Wang and Hao Xiang and Xinchen Xie and Shuhao Xing and Xiaoyu Xing and Wanghan Xu and Xinyu Yang and Yajie Yang and Chengfeng Zhao and Haoran Zhao and Ruojun Zhou and Yunhua Zhou and Yicheng Zou and Kun Cai and Qiye Cai and Xinmeng Che and Haodong Chen and Jiabei Chen and Jiahao Chen and Jiayi Chen and Yujia Chen and Lizhi Cui and Youheng Dai and Xin Deng and Yi Dong and Shihan Dou and Chenya Gu and Xu Guo and Ding Han and Feiyang Hao and Haotan He and Jie Hou and Binze Hu and Zijian Hu and Junhao Huang and Huicheng Jiang and Jiazhen Jiang and Shufan Jiang and Jiahao Kuang and Bowen Lai and Bo Li and Jiaqiang Li and Peng Li and Qilong Li and Zhuoqun Li and Jiaxiang Liu and Shuainan Liu and Tong Liu and Yi Liu and Zhonghang Lu and Jianwen Luo and Yanyi Luo and Huijie Lv and Ningsheng Ma and Zerun Ma and Houcheng Min and Chengjun Pan and Qiyuan Peng and Xiaoxuan Peng and Jianmin Qian and Jiantao Qiu and Wanying Ren and Huayu Sha and Jifei Shan and Zixin Shang and Bing Shao and Zhuohui Sheng and Jiayang Shi and Yang Shu and Aierpanjiang Simayi and Sirui Song and Yuxiao Song and Zhe Sun and Zhichao Sun and Wenzhe Tan and Wenhui Tian and Zhongbo Tian and Hanchen Wang and Pengbo Wang and Rui Wang and Yiding Wang and Yuhui Wang and Zhiheng Xi and Caijun Xu and Chao Xu and Yongfeng Xu and Xiaolei Yang and Zhixiong Yang and Qian Yao and Shihong Yi and Yuankai Ying and Jia Yu and Dingbo Yuan and Hao Yuan and Junjie Yuan and Bo Zhang and Caixian Zhang and Qiuyinzhe Zhang and Jiyuan Zhao and Penghao Zhao and Ying Zhao and Pujun Zheng and Xiaoxue Zhong and Xiaohao Zhou and Xinyu Zhou and Dongsheng Zhu and Guanru Zhu and Yulun Zhu and Yaojie Lu and Tao Ji and Hongyu Lin and Yutao Zhu and Pengfei Cao and Guoxiu He and Xianpei Han and Ben He and Zhicheng Dou and Kang Liu and Qi Zhang and Le Sun and Jun Zhao and Ji-Rong Wen and Xuanjing Huang and Yu-Gang Jiang and Bowen Zhou},
      year={2026},
      eprint={2609.15818},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.15818}, 
}

_{Built by the Shanghai Artificial Intelligence Laboratory.}