EASYHUB JOURNAL / 2025-11
AI news in November 2025
8 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
Fara-7B experiments with screen-grounded web actions
Fara-7B predicts web actions from screenshots and targets local computer use. Its experimental release emphasizes sandboxing and confirmation at sensitive steps; on-device execution alone does not ensure safety or reliability.
- IT之家
HunyuanOCR uses a 1B end-to-end model for text and documents
HunyuanOCR is a 1B end-to-end text and document model for inputs such as documents, handwriting, receipts and screenshots. Tencent opens code, weights and an online demo, combining recognition and complex document handling at a small scale. Reported benchmark results do not replace verification on real documents.
- IT之家
MiMo-Embodied studies shared representations for robots and driving
MiMo-Embodied opens model assets for shared perception and planning across embodied domains. Research evaluations are not driving-safety certification or a substitute for validating a complete robotics system.
- Tencent
HunyuanVideo 1.5 pairs an 8.3B generator with video upscaling
HunyuanVideo 1.5 opens inference code and weights for an 8.3B generator covering text- and image-conditioned clips. A separate super-resolution stage improves output resolution, allowing creators to organize generation and upscaling as modular steps rather than expecting a single model to perform every part of video production.
- IT之家
VibeThinker-1.5B explores compact mathematical and coding reasoning
Weibo AI opens a 1.5B reasoning model with a focus on training methods and mathematics and coding evaluations. Such results do not establish general superiority in conversation, knowledge coverage or complex tool use.
- Google
NotebookLM Deep Research brings reports and sources into the same notebook
NotebookLM adds Deep Research to plan web searches, assemble reports and import both the report and its sources into a notebook. It also expands support for Sheets, Word documents, Drive PDFs and images. Availability is staged, with image support arriving after the other announced source types.
- IT之家
Omnilingual ASR opens multilingual recognition models and speech data
Meta releases multiple ASR sizes and data for underrepresented languages. Model and dataset licenses differ, and broader language coverage does not imply identical accuracy across accents and recording conditions.
- IT之家
UCM opens multi-level KV-cache and inference-memory management
UCM coordinates inference frameworks, compute and storage around longer-context caches. Maximum throughput and latency improvements come from particular configurations and should not be combined into a universal deployment promise.