As of Beijing time October 23, 2026, Hugging Face repository internlm/Atria-Dawn shows commits from 2026-10-23T03:21:40Z onward for the formal Instruct build. An FP8-quantized Instruct sibling is at internlm/Atria-Dawn-FP8. Both formal listings show a 512K context window; the Preview and Preview FP8 entries remain listed at 256K.
The model card introduces Atria Dawn as an agentic model from Shanghai Artificial Intelligence Laboratory for research and engineering work that needs ongoing tool use and multi-step delivery. The card puts the MoE foundation GLM-5.2 at about 744B parameters; Hugging Face’s safetensors count shows about 753B—two published figures from different places. This is Shanghai AI Lab opening formal weights, not Zhipu announcing a new GLM generation.
Inputs are text only. The card’s Codex setup notes that attaching images can return `400` with a message that Atria-Dawn is not multimodal—so client configs need a text-only modality declaration if you wire it into Codex. Serving notes on the card point to SGLang and vLLM for local deploy, and the same page links regional API consoles for hosted access.
The BF16 weight shards on the main repo sum to about 1.51TB on disk; the FP8 repo’s safetensors sum to about 755.7GB. Those numbers are download and storage size, not a stated minimum GPU memory to run the model. The model card states the code and weights are released under the MIT License. Related reading: Sage-A1.5 Open Weights: Self-Verification Is Opt-In and Off by Default, MiniMax M3.1 Flash Open Weights Are on Hugging Face, and Recursive Releases Flint, an Open-Weights Coding Model Built to Run on a Laptop.
As for end-to-end local deploy cost, actual VRAM needs, API pricing by region, and how the formal weights behave on your workloads—we don’t have that information yet. If you’ve tried them, you can send feedback through the site-owner contact link in the footer. Opening weights is not the same as a free hosted API.