Papers · Models · Repos · Every region
Up to three papers, models or repos a day that builders should know about, each linked to its primary page, with one line on why it made the cut. Picked by our scan and our rubric, checked and approved by a human editor.
Picked because: 27B-parameter vision-language model built on Qwen3.5, with a native 262,144-token context extendable to about 1 million; Qwen's own coding and agent benchmark scores are on the card. The card lists an Apache-2.0 licence.
From States Take AI In-House: Abu Dhabi, Russia, South Korea
Picked because: Submitted 5 November; a one-step diffusion fix for artifacts in 3D head avatars, with code linked at the authors' project page. The abstract page shows no affiliations; the region is our inference from the scan's KAIST tag.
From States Take AI In-House: Abu Dhabi, Russia, South Korea
Picked because: Submitted 5 November; per its authors, a linear-attention design that cuts perplexity 1.8 to 2.0% against a gated delta-rule baseline from 166M to 1.34B parameters. The abstract page shows no affiliations; the region is our inference from the scan's Google tag.
From States Take AI In-House: Abu Dhabi, Russia, South Korea
Picked because: 280B-parameter mixture-of-experts base checkpoint with 14B active per token, released 5 November before any instruction tuning. The custom licence covers use only outside the EU, the UK and South Korea, and asks for "Powered by ERNIE" on commercial products and a notice or authorisation by revenue (US$50 million). Benchmarks are Baidu's own.
Picked because: Submitted 5 November; adds 3D spatial alignment to robot-policy pre-training, starting from a pretrained pi05 policy, with the extra parts used only in training so the inference design is unchanged. The abstract page shows no affiliations; the region is our inference from the HKUSTGZ organisation tag on Hugging Face.
Picked because: Submitted 4 November; an encoder-free robot-control model that, per its authors, beats encoder-based designs under controlled comparisons. The abstract page shows no affiliations; the region is our inference from the mmlab-ntu organisation tag on Hugging Face.
Picked because: 3B-parameter image-text-to-text model for very long documents; Baidu released it on 27 October and published the paper on 2 November. The model page lists the MIT licence.
From Schools Rein In AI: Australia Limits It, Google Ties Search
Picked because: Reasons over continuous hidden states rather than text tokens before generating images; submitted 30 October. Authors' affiliations are not shown on the abstract page, so the country is our inference from author names.
From Schools Rein In AI: Australia Limits It, Google Ties Search
Picked because: Kling Team report arguing video quality depends on multimodal understanding across the data, conditioning, alignment and evaluation chain; submitted 3 November. Affiliation is not stated on the abstract page.
From Schools Rein In AI: Australia Limits It, Google Ties Search
Download: JSON · CSV (free to reuse with a link to AI Worldwire). How we pick.