We’re in San Diego this week for #NeurIPS2025! Stop by the Meta booth (#1223) to meet our team and check out: Demos of our latest research including DINOv3 and UMA Lightning talks from researchers behind SAM 3, Omnilingual ASR and more (see schedule below) Hands-on
RESEARCH
-

Robotic hands and humanoid developments dominate robotics news
By
–
Top stories in robotics today: – Lifelike robotic hand mimics human grip
– Hugging Face’s robotics lead launches startup
– China sounds alarm on humanoid bubble
– Robots rebuild Pompeii’s shattered frescoes
– Quick hits on other robotics news Read more: https://
robotnews.therundown.ai/p/the-creepies
t-robot-just-got-hands
… -

Baidu ERNIE 5.0 Surpasses GPT-5 on Key Benchmarks
By
–
6️⃣ Baidu introduced ERNIE 5.0, a multi-modal model that reportedly surpasses GPT-5 on key benchmarks across text, image, audio, and video. pic.twitter.com/xwR2GtXkBx
— Futurepedia – Learn to Leverage AI (@futurepedia_io) 1 décembre 2025Baidu introduced ERNIE 5.0, a multi-modal model that reportedly surpasses GPT-5 on key benchmarks across text, image, audio, and video.
-

DeepSeek Math V2: Open Weights Model for Mathematical Problem Solving
By
–
DeepSeek Math V2
A huge open weights math model aimed at proofs and problem solving. It is tuned for contests like IMO and Putnam and licensed so companies can use and fine tune it for real world scientific and engineering work. -
Qwen3-VL: Blueprint for Next-Gen Visual AI
By
–
If you want to understand where multimodal AI is actually heading, this report is required reading. Qwen3-VL is the clearest blueprint for next-generation visual reasoning systems.
-
Data Pipeline Powers Real-World VLM
By
–
What ties everything together is the data pipeline. They rebuilt OCR, parsing, and multilingual data from scratch, combining synthetic HTML, pseudo-labels, and region-aware markup all feeding a 256K-token multimodal engine. This is how you build a real-world VLM, not a
-

Advanced Long-Document Understanding Capabilities
By
–
Long-document understanding is where this model feels unfair. It parses multi-page PDFs, reconstructs layout, fuses visual regions, and aligns text+image reasoning across hundreds of pages in one context window.
-

Qwen3-VL excels in multilingual OCR
By
–
The multilingual OCR is insane. Most VLMs fall apart outside English and Chinese. Qwen3-VL is scoring above 70% across 32 different languages… even on messy real-world photos. This is the part that makes it usable globally, not just in benchmark labs.
-

Qwen3-VL Redefines Vision-Language Models
By
–
Holy shit… Qwen3-VL just rewrote the rules for vision-language models This thing doesn’t behave like a “VL model.” It behaves like a full-stack multimodal machine that can read images, reason through them, parse dense text, understand diagrams, and generate step-by-step
-

NYT Headlines and AI Coverage: Questioning Media Narratives
By
–
There are lies, damned lies, and NYT headlines.
