EASYHUB JOURNAL / 2025-07
AI news in July 2025
17 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
Volcano Engine announces open Coze capabilities and model services
Volcengine announces access to Coze’s core capabilities, enterprise model hosting and a Responses API for application workflows. The same event includes SeedEdit 3.0 image editing and interpretation updates. Image editing is offered through the platform API and consumer apps; these are separate services rather than one model.
- IT之家
NotebookLM Video Overviews explain sources through narrated slides
NotebookLM adds Video Overviews that organize supplied documents, charts and images into narrated slides, alongside Studio changes. English was supported at launch, with other languages planned. The feature explains source material rather than acting as a general-purpose model for generating live-action footage.
- IT之家
Wan 2.2 expands video generation with a unified 5B variant
Wan 2.2 adds text- and image-conditioned models plus a smaller unified variant, with finer visual controls. Parameter counts and memory requirements differ by model and deployment setup.
- IT之家
Hunyuan World 1.0 opens editable 3D scene generation from text or images
Hunyuan World 1.0 combines panoramic generation with layered reconstruction to turn text or images into navigable 3D scenes. Its editable mesh assets allow changes to foreground objects, ground and sky and can enter Unity, Unreal Engine or Blender workflows, rather than remaining only a generated video.
- IT之家
WPS Lingxi brings conversational creation into Office workflows
Kingsoft introduces WPS AI 3.0 and Lingxi at WAIC, bringing multi-turn instructions into document creation, presentations and voice interaction. A knowledge-base product accompanies the release. The focus is an integrated Office workflow rather than a separately launched general-purpose foundation model.
- IT之家
Jiutian 3.0 shares 8B-scale models for data, mathematics and code
China Mobile presents task-specific models and evaluation resources across structured data, mathematics and code. The smaller checkpoints and accompanying benchmarks support domain-focused experimentation rather than only general chatbot use.
- IT之家
Google Opal tests visual, editable workflows for small AI apps
Opal enters testing in the US through Google Labs, turning requests into small web applications with editable input, model and output steps. Users can inspect instructions, add steps and share applications for others to try with a Google account. The initial rollout was limited rather than globally available.
- The Verge
Figma Make expands access while publishing remains seat-dependent
Figma Make expands prompt-to-prototype access beyond full-seat subscribers, allowing more users to experiment in drafts and refine interactions. At the announcement, public publishing still required a Full seat on a paid plan. Broader trial access did not mean unrestricted free publication.
- IT之家
Qwen3-Coder and Qwen Code expand open coding-agent workflows
Qwen releases an agent-oriented coding model alongside Qwen Code tooling, emphasizing context and tool interaction. Code quality, browser-task behavior and serving requirements are distinct parts of evaluating the workflow.
- JetBrains
Junie brings asynchronous coding to GitHub and adds MCP support
Junie adds asynchronous GitHub tasks alongside its IDE workflow, plus MCP connections to external tools and remote-development support. Developers can delegate work and review changes afterward. In this announcement, the GitHub early-access program is limited to JVM and PHP, while remote development initially covers macOS and Linux.
- IT之家
Kiro’s preview organizes AI coding around specifications and tasks
AWS introduces Kiro in preview, organizing development around requirements, system design and implementation tasks before code changes. It emphasizes inspectable specifications and maintained project documentation. The free preview was a launch-time arrangement, not a promise of permanent free access.
- IT之家
Kimi K2 releases base and instruction variants for coding and agents
Kimi K2 launches with base and instruction checkpoints plus API access for coding and agent tasks. Its active MoE parameter count is not the same as stored model size, so deployment resources and license conditions still matter.
- IT之家
Comet’s initial release brings page Q&A and task assistance to browsing
Perplexity launches Comet on Chromium, adding a sidebar for page understanding, summaries and task assistance, plus Chrome-data import. Its initial Mac and Windows rollout used invitations or Max access. Actions involving email, purchases or other accounts still need appropriate authorization and review.
- Liquid AI
LFM2 releases 350M, 700M and 1.2B models for on-device workloads
Liquid AI releases dense LFM2 models at 350M, 700M and 1.2B parameters for low-latency on-device workloads. Downloadable weights use a license with a commercial revenue threshold at launch, not unrestricted Apache 2.0 terms. Reported speed comparisons are tied to the publisher's test setup.
- 证券时报
Feishu updates enterprise knowledge, meetings and app-building tools
Feishu updates knowledge Q&A, AI meetings, Aily and Miaoda in one product event. Answers use messages and documents the employee is permitted to access; meeting tools organize notes and actions, while app builders support business workflows. Connecting enterprise information and respecting access permissions remain prerequisites.
- Hugging Face
SmolLM3 opens a 3B model with selectable reasoning modes
SmolLM3 releases 3B base and instruction checkpoints with thinking and non-thinking modes, six-language support and an openly described training recipe. Context training reaches 64K, with YaRN extending inference to 128K. Developers can choose a reasoning mode according to task complexity and latency needs.
- IT之家
AniSora V3 previews animation-focused video generation
AniSora V3 targets animated scenes, comic adaptations and virtual-character content with improvements to motion and style. The preview supplies material to evaluate and edit, rather than removing the need for post-production.