Discover LongVT on Hugging Face! Paper: https://t.co/ZE4FjbRNxB Models & Data Collection: https://t.co/3uiMckxIgk Try the demo:

LongVT: Incentivizing "Thinking with Long Videos" This new framework for LMMs uses Multimodal Chain-of-Tool-Thought, enabling global-to-local reasoning in long videos via native video cropping to tackle hallucinations and outperform baselines.

Z-image's paper is out on @HuggingPapers 🔥

Dive into self-verifiable math reasoning! DeepSeekMath-V2 achieves a near-perfect 118/120 on Putnam 2024 by using an LLM-based verifier to iteratively improve proofs. Paper: https://t.co/pP5WiYRSC0 Model:

DeepSeek-AI just dropped DeepSeekMath-V2 on Hugging Face This new self-verifying mathematical reasoning model generates and validates its own proofs, achieving gold-level scores on Olympiad-level benchmarks like IMO 2025 and CMO 2024.

This groundbreaking model allows new languages to be added with few examples. It achieves SOTA performance, making speech tech more inclusive for communities worldwide. Try the demo: https://t.co/JMDs8ebEqF Get the model:

Meta just released Omnilingual ASR on Hugging Face It's an open-source speech recognition system supporting over 1600 languages, including hundreds never before covered by ASR technology.

Experience the next generation of conditional AI models with ViBT! 🚀 Try the interactive demo: https://t.co/CICxVpqQ01 🧠 Grab the 20B model: https://t.co/5vwjsgN9Qs 📄 Read the full paper:

ViBT: The First Vision Bridge Transformer at 20B Parameters This groundbreaking framework pioneers data-to-data translation, directly modeling trajectories for conditional image & video generation. It's incredibly efficient, up to 4x faster, handling complex tasks with ease.

DeepSeek-V3.2 introduces Sparse Attention for long-context tasks and excels in agentic tool-use. It achieved gold medals in the 2025 IMO & IOI. Explore the model: https://t.co/bYOxe3y5xc Read the technical report:

DeepSeek AI just unveiled DeepSeek-V3.2 on Hugging Face This new model harmonizes high computational efficiency with superior reasoning and agent performance, with a special variant surpassing GPT-5.

Project page: https://t.co/kRLi5jpd7z Code: https://t.co/CRcaAOsbzm Read the full paper on Hugging Face:

Architecture Decoupling is Not All You Need Researchers explore why decoupling unified multimodal models works. Their key insight: it pushes models toward task-specific interaction patterns. They propose Attention Interaction Alignment (AIA), a new loss to get similar benefits

This looks to be this paper!

Apple just released Starflow on Hugging Face. A brand new model from the tech giant, opening doors to unexplored possibilities in AI development. https://t.co/uj9HrreXRV

Dive into the details of AnyTalker! This breakthrough in multi-person video generation offers unmatched lip sync, visual quality, & natural interactivity across diverse inputs. Paper: https://t.co/UBO7bNVdX2 Model (1.3B):

Say hello to AnyTalker for dynamic multi-person videos from HKUST-C4G & Video Rebirth! Unveiling an innovative framework for scalable multi-person talking video generation. AnyTalker drives arbitrary identities with natural interactivity, all while being data-efficient.

Experience a thinking-editing-reflection loop for superior image quality & instruction-following. ReasonEdit-S (based on Step1X-Edit) boosts ImgEdit by +4.3%, GEdit by +4.7%, and Kris by +8.2%! Get the model & try the demo: Model: https://t.co/qGDCTfhhy2 Demo:

REASONEDIT: Towards Reasoning-Enhanced Image Editing Models just dropped! This new framework leverages MLLM thinking & reflection to interpret abstract instructions and iteratively refine image edits, pushing the boundaries of what's possible in generative AI.

Experience Z-Image-Turbo: sub-second inference (8 steps!) on <16GB VRAM. Get the paper, model & demo on @HuggingFace: 📄 Paper: https://t.co/clGPAj0Ygh 💾 Model: https://t.co/kQcMCrBywa ✨ Demo: