Shanghai AI Lab is the Hub story today. internlm updated Atria Dawn Preview, an MIT 744B mixture of experts on GLM-5.2, and the matching paper is #1 on Daily Papers. Intern-S2-397B is the non-Preview scientific multimodal repo from September 13. Hugging Face PEFT absorbed ShadowPEFT as a first-class method. Daily Papers also carries Grouped Value Attention, a compact KV-cache design with no Hub weights yet.
On the Hub
Atria Dawn Preview: Shanghai AI Lab MIT 744B MoE
Shanghai AI Lab posted internlm/Atria-Dawn-Preview on September 11 and updated the card today. The preview is a 744B mixture of experts built on GLM-5.2. It is text-only instruct, with a 256K context window on the Hub table, under MIT. The BF16 pack is about 1.5 TB of shards. An FP8 sibling was updated today as well.1
Paper arXiv 2609.15818 was submitted to Hugging Face Daily Papers on September 15 and is #1 Paper of the day. The paper claims the highest reported score on five of sixteen benchmarks. It includes a case study of 769 tasks from 56 participants.2
Card-claimed scores on the rows the model card bolds as best-in-row are AutomationBench 53.8, BFCL v4 77.0, CyberGym 86.5, DeepSearchQA 96.0, and BrowseComp 92.5. SWE-bench Pro is 59.6 on the same table, which is not the row leader. Serve notes point at SGLang v0.5.13.post1+ and vLLM v0.23.0+ using the GLM-5.2 guides.
ShadowPEFT: first-class method in PEFT main
A community article published September 15 says ShadowPEFT is now a first-class method in the Hugging Face PEFT library. The paper is arXiv 2604.19254 from April. Install is from source with git+https://github.com/huggingface/peft.git. The stable pip release is still v0.20.0.3
You wrap a base model with ShadowConfig and get_peft_model, the same entry point as LoRA. The default shadow is a mirror network at rank 8. Merge is not supported; detach with model.base_model.unload_shadow(). The shadow rides the decoder with its own prediction head, and you can run the detached shadow alone.4
Intern-S2-397B: Shanghai AI Lab Apache-2.0 scientific multimodal
The repo internlm/Intern-S2-397B landed on September 13 under Apache-2.0. The card calls it a multimodal foundation model for scientific intelligence and long-horizon agents. Pipeline tag is image-text-to-text. Hub safetensors total 403,423,094,768 parameters; the repo is named 397B. Weights are about 807 GB.5
Config is Qwen3_5MoeForConditionalGeneration with 60 text layers, 512 experts, and ten experts per token. A vision tower is present. Thinking mode is on by default. Serve options on the card are LMDeploy, vLLM, and SGLang.
Grouped Value Attention: compact KV cache paper
arXiv 2609.13285 was submitted to Hugging Face Daily Papers on September 15. It was published September 8. Authors are Vishesh Tripathi, Abhay Kumar, and Ramsha Khan.6
The method stores grouped values and reconstructs content keys with a learned linear map that can be absorbed into the query at decode. A small shared decoupled-RoPE channel keeps a positional key. The paper claims persistent cache scalars are about 45 to 47 percent below matched GQA.
At 350M scale on 30B FineWeb-Edu tokens, the 16-d positional variant scores 44.18 average on five tasks, against 44.36 for GQA and 43.88 for MLA. No Hub weights are present. The paper notes custom decoding kernels, with an open source release planned soon.
