internlm/Atria-Dawn-Preview

中文 ｜ English

Atria Dawn Preview: From Research Questions To Verifiable Results

---

🖥️Website 💬GitHub

Introduction

Atria Dawn Preview is a preview version 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 is designed for research and engineering scenarios that require continuous environmental understanding, tool use, and multi-step task completion. The model helps users drive open-ended problems toward executable, verifiable, and reproducible results. It can support problem analysis, solution design, tool use, code implementation, experiment execution, result analysis, and failure recovery by combining task objectives with environmental feedback.

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:

• Discovery: Retrieving and organizing evidence, conducting deep research, and turning research questions into executable experimental plans.
• Creation: Building software, interactive applications, games, data visualizations, and machine learning systems.
• Delivery: Transforming documents, data, and design requirements into reports, presentations, and other structured deliverables.
• Cybersecurity: Analyzing security issues, validating vulnerabilities, applying fixes, and performing re-validation in authorized environments.

Model Downloads

Model	Description	Context	Hugging Face	ModelScope
Atria-Dawn-Preview	Instruct model	256K	Model	Model
Atria-Dawn-Preview-FP8	FP8-quantized Instruct model	256K	Model	Model

Evaluation Results

We conducted a comprehensive evaluation of Atria Dawn Preview. The table below presents benchmark results.

Benchmark	Atria Dawn Preview	DeepSeek V4 Pro 0813	KIMI K3	Qwen 3.8 Max	GLM 5.3	GPT 5.6 sol	Claude Opus 5
AutomationBench	53.8	41.7	45.9	49.7	49.2	45.7	49.4
BFCL v4	77.0	71.4	69.1	-	74.1	-	-
CyberGym	86.5	83.3	78.7	73.8	84.5	83.6	-
DeepSearchQA	96.0	-	95.9	-	94.7	93.2	-
Workspace-Bench-Lite	68.2	58.1	65.8	67.4	67.7	60.5	70.1
BrowseComp	92.5	83.4	91.2	-	-	92.2	90.8
SkillsBench	66.4	65.0	51.9	66.7	63.3	62.5	63.7
Workspace-Bench	65.0	55.7	60.6	63.9	63.9	56.0	65.8
MLE-bench Lite	86.2	86.8	85.8	81.3	80.8	88.9	88.0
WideSearch	81.9	-	79.6	81.9	82.7	83.3	-
DeepResearch Bench II	51.1	46.6	51.3	49.2	52.7	50.7	54.1
τ³-Bench Banking	41.2	44.3	37.1	55.2	40.2	46.9	48.7
Terminal-Bench 2.1	78.3	78.7	-	89.3	85.4	85.1	90.2
GDPval	1583	1517	1611	1722	1667	1682	1768
SWE-bench Pro	59.6	58.3	61.6	65.1	60.3	61.4	74.7
JobBench	50.3	54.1	54.3	52.7	58.2	45.4	68.0

Deployment & Online Access

Atria Dawn Preview supports both local deployment and hosted access. For online access, use the service corresponding to your region.

Region	Access	Tutorial
International	Link	Tutorial
China	Link	Tutorial

For local deployment, please refer to the following deployment guides.

• SGLang (v0.5.13.post1+) — see cookbook
• vLLM (v0.23.0+) — see recipes

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": []
      }
    ]
}

License

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

Contact

For questions or suggestions, please contact us by email or through GitHub and other project platforms.