EASYHUB JOURNAL / 2025-10
AI news in October 2025
9 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
ima 2.0 announces task-mode testing for source-based reports and podcasts
ima 2.0 announces task-mode testing beginning October 24. Users can attach libraries, web pages, files or notes and have an agent plan a report or podcast, with selectable speakers and voices. Library summaries and parallel topic tasks are added; the announcement is not a general-availability release.
- IT之家
Qwen3-VL adds 2B and 32B dense visual-language variants
The new dense variants broaden device and server choices within Qwen3-VL, with instruction and thinking modes. Precision, image resolution and context length remain important parts of their practical resource requirements.
- IT之家
WorldMirror opens multi-view and video-driven 3D reconstruction
WorldMirror extends Hunyuan World into multi-view and video-driven 3D reconstruction, optionally using camera and depth information to predict point clouds, depth, surface normals and new views. Code, weights and an online 3DGS preview are available. Its focus is scene geometry recovery, not a complete interactive game.
- IT之家
DeepSeek-OCR studies visual compression of document context
DeepSeek-OCR combines visual encoding and language decoding to represent long documents with fewer visual tokens. Compression and recognition trade off against each other, so test accuracy cannot be assumed for every receipt or table.
- IT之家
Haiku 4.5 targets responsive assistants and coding workflows
Haiku 4.5 emphasizes responsiveness and cost for assistants and coding. Launch prices and comparisons describe their original context, and undisclosed model size should not be replaced with an invented parameter count.
- IT之家
Manus 1.5 combines full-stack applications, collaboration and a library
The update adds application backends, databases and shared project workspaces. Vendor speed comparisons reflect its task measurements, while generated applications still require security, data and business-logic testing.
- IT之家
Youtu-Embedding opens a 2B text representation model and training framework
Tencent Youtu opens a 2B embedding model that converts text into comparable semantic vectors for retrieval, similarity, classification, clustering and knowledge-base applications. Weights, inference code and a training framework are released. It is a retrieval component for systems such as RAG, not a chatbot generating final answers.
- IT之家
SINQ opens a lower-overhead model quantization method
The report examines a quantization method from Huawei’s Zurich team aimed at lowering memory and calibration overhead. Its hardware comparisons are workload-specific, not a general equivalence between consumer and enterprise GPUs.
- The Verge
Jules adds CLI and API access for coding-task integrations
Jules adds terminal and API access beyond its existing interfaces. Workspace support was still scheduled for later that month, so the report distinguishes new access points from planned integrations.