Kandinsky Lab's 6.0 Video is Hugging Face's top Daily Paper on 6 October. The same morning the lab tagged the MIT stack open-source on GitHub and put a Distill Space on the Hub. Abu Dhabi's Falcon team posted a dialect chat model and a 270-million-parameter Arabic OCR, both with demos you can try and neither with a new public weight card.
On the Hub
Kandinsky 6.0 Video: MIT synced audio-video, paper and Space today
Kandinsky Lab's 6.0 Video is Hugging Face Daily Paper number one for 6 October (arXiv 2610.05608).1 The same morning the GitHub README tagged the stack open-source, and a Distill-5s Space went up. The family is Video Lite at 3 billion parameters, Video Pro at 29 billion, and a 1.4 billion super-resolution head, all under MIT.
It is a dual-stream CrossDiT: a pretrained video stream plus a newly trained audio stream, then fine-tuning, RL, and a 10-step distillation. Lite and Pro generate 5-second clips from text or an image, with 44 kHz audio and lip-sync. The super-resolution head scales 2x, 2.25x, and 4x, up to 1920 by 1080.
The paper's submitter says RL post-training cuts Pro speech word-error rate by 47 percent, and that VABench has Pro ahead of LTX 2.5 on speech quality, AV alignment, lip-sync, and visual realism. Diffusers weight cards have been on the Hub since 9 September. What changed this morning is the paper, the open-source tag, and the demo.
Non-distilled Pro is slow even on a 5090. The README times a 5-second clip at 754 seconds after warmup, so the distilled 10-step path is the one you actually run.
TII Falcon-Emirati-7B: dialect chat, no Hub card yet
The Technology Innovation Institute posted a Hugging Face team article this morning about Falcon-Emirati-7B.2 It is a 7-billion-parameter dialect specialist on Falcon-H1-Arabic, aimed at Emirati Arabic (vocabulary, nabati poetry, proverbs) rather than textbook Modern Standard Arabic. The base is a hybrid of Mamba SSM and Transformer attention in every block.
TII reports 84.83 percent on Alyah (1,173 native multiple-choice items) and 85.57 percent on ArabCulture-Dialogue UAE, ahead of ALLaM-7B, Jais-2-8B, and Fanar-2-27B on that culture set. On open-ended Alyah, with Gemini 3.7 Flash as judge, they say it is the only model that consistently answers in Emirati rather than Modern Standard Arabic. Those numbers are TII's own benches.
The post points readers to the Falcon Chat app. There is no public weight card on the Hub, no license stated, and no GGUF to download.
TII Falcon-OCR-Arabic: 270M on Arabic pages, playground today
Same morning, TII published a community blog on Falcon-OCR-Arabic, a 270-million-parameter early-fusion OCR they say is second of 17 on Arabic text and first on tables.3 A prompt category selects plain text, LaTeX, or HTML tables. The playground is up.
On TII's Arabic document bench (11,974 real pages, 15 categories, diacritics counted), they report 81.87 percent text accuracy and 59.95 percent Table TEDS. That table score is first of 17 models they plotted, ahead of Gemini 3.5 Flash at 51.30 percent TEDS even though Gemini leads on raw text at 84.34 percent. English Falcon-OCR, the February Hub card, scores 55.39 percent text and 24.83 percent TEDS on the same Arabic set.
The Hub still lists tiiuae/Falcon-OCR, created 22 February. There is no new public Falcon-OCR-Arabic weight card. The playground is what you can try today.
