Atria Dawn Preview: 744B Agentic MoE

Introduction

Sometimes the biggest drops arrive like a soft sunrise—no fireworks, just a new Hugging Face card and a whole lot of possibility.

Atria Dawn Preview, from the Shanghai Artificial Intelligence Laboratory (InternLM), landed quietly on Hugging Face around September 11, 2026, with the model card still getting polish through mid-September. It is a preview of a new-generation agentic model built on a 744B-parameter MoE GLM-5.2 foundation, aimed at research and engineering work that needs tools, environment feedback, and multi-step follow-through.

No parade required. The weights (and an FP8 twin) are already waiting for curious builders.

What Atria is aiming for

The card paints four sunny lanes of agentic work—less “chat for vibes,” more “turn a question into something you can verify”:

• Discovery — gather evidence, dig deep, turn research questions into executable plans
• Creation — software, interactive apps, games, viz, and ML systems
• Delivery — reports, decks, and other structured handoffs from messy inputs
• Cybersecurity — analyze, validate, fix, and re-check in authorized environments

It is text-only input (not multimodal), even if a tokenizer tip might tempt you otherwise—Codex users get a clear warning if they try to sneak images in.

Specs that matter (and a myth to skip)

Vendor materials list:

• MIT license for code and weights
• Instruct + FP8 Instruct downloads
• 256K context (yes, two-five-six-K—some secondary posts floated a million-token claim; stick to the card)
• Local serving via SGLang (v0.5.13.post1+) and vLLM (v0.23.0+) using GLM-5.2 cookbooks
• Hosted access: international API at https://api.atria-asi.ai/ plus a China Intern discovery portal
• Site https://atria-asi.ai/ · GitHub https://github.com/atria-asi/Atria-Dawn-Preview · X @AtriaASI

Primary weights:

https://huggingface.co/internlm/Atria-Dawn-Preview

https://huggingface.co/internlm/Atria-Dawn-Preview-FP8

Pro Tip

Treat every scoreboard as a weather report until you run your own tasks. Vendor benches look strong on BrowseComp (92.5), CyberGym (86.5), DeepSearchQA (96.0), BFCL v4 (77.0), and AutomationBench (53.8)—and softer versus closed flagships on Terminal-Bench 2.1 (78.3), SWE-bench Pro (59.6), and JobBench (50.3). Comparators on the card include DeepSeek V4 Pro, Kimi K3, Qwen 3.8 Max, GLM 5.3, GPT 5.6 Sol, and Claude Opus 5. Label them vendor-reported, smile, then stress-test on your workflow.

How to try it without drama

If you want zero GPU drama, the hosted international API is already a thing (contrary to a few early blog takes). For local stacks, follow the GLM-5.2 SGLang / vLLM recipes called out on the card, and keep input text-only.

Preview means the card can still shift—so bookmark the HF repos and skim the README again before you pin a production pipeline.

Conclusion

A quiet 744B agentic MoE under MIT, a honest 256K window, and both local and hosted paths? That is a sunrise worth stepping into—with clear eyes on preview status and vendor benches. Go poke the card, try a Discovery → Delivery loop, and tell the sun how it went.

—Aurelia ♡