EASYHUB JOURNAL / 2025-09
AI news in September 2025
15 stories, ordered by their original dates. Open a title for its summary, image and source link.
- IT之家
Hunyuan Image 3.0 opens an 80B model for detailed prompts and text rendering
Tencent opens HunyuanImage 3.0, an 80B native multimodal image-generation model emphasizing knowledge-grounded understanding of detailed prompts and longer text rendering. It brings subject descriptions, composition requirements and written content into one generation task for more precisely specified imagery.
- IT之家
Hunyuan 3D-Omni and 3D-Part add geometry controls and component generation
Tencent releases two complementary 3D tools together. 3D-Omni uses multiple conditions to control geometry and pose, while 3D-Part combines P3-SAM segmentation with X-Part generation to produce separate components. Weights and inference code are opened, and component tools reach 3D Studio for asset editing and print preparation.
- IT之家
Kimi OK Computer enters limited testing for virtual-computer work
Kimi introduces OK Computer, a K2-based mode using its own virtual computer for websites, data analysis, media generation and presentations. It plans tasks, invokes tools and delivers outputs from a user’s goal. The launch is a limited rollout, initially prioritizing users who had previously supported Kimi.
- 新智元
Code World Model studies execution prediction for coding
CWM uses execution-related training to explore predicting state changes during coding. It is a research direction, and a simulated execution remains different from running real tests against a proposed patch.
- IT之家
AgiBot GO-1 opens embodied-model assets and development workflows
GO-1 combines embodied-model assets with data, training and deployment workflows. Cross-robot tests support portability, while safety, control interfaces and new task data remain deployment-specific responsibilities.
- ComfyOrg
ComfyUI adds native Wan animation and Qwen multi-image editing workflows
ComfyUI adds native workflows for Wan2.2 Animate character animation and replacement, plus Qwen-Image-Edit-2509 composition from one to three images. The announcement targets ComfyUI 0.3.60 with separately downloaded models; desktop-package support is still forthcoming at that point, rather than available through every distribution immediately.
- IT之家
MobileLLM-R1 targets reasoning tasks with sub-billion models
MobileLLM-R1 offers sub-billion variants focused on mathematics, code and science. This specialization provides options for constrained devices, but it should not be treated as equivalent to a general conversational assistant.
- IT之家
Granite-Docling-258M preserves document structure through DocTags
IBM’s 258M-parameter Granite-Docling represents document content, tables, formulas and reading structure in DocTags for downstream conversion. It supports document-processing workflows rather than a general chat interface. Multilingual capability, including Chinese, still had maturity limitations at the report date and needs application-specific checks.
- IT之家
MiMo-Audio opens research into end-to-end speech pretraining
Xiaomi shares speech-model and encoding components to study few-shot task generalization from audio pretraining. Tokenizers and complete speech systems are different components, so their parameter counts must be identified separately.
- The Verge
Notion Agent links pages, databases and search in task workflows
Notion’s agent can plan work across pages, databases and connected information sources while retaining editable preferences. Fully automated custom agents were still planned, and access remains bounded by workspace permissions.
- IT之家
Tongyi DeepResearch opens models and workflows for research agents
Tongyi DeepResearch opens components linking search, reading and multi-step answering. An inspectable pipeline helps evaluation, while retrieved evidence, citations and conclusions still need verification.
- Replit
Replit Agent 3 adds browser testing and automation building
Replit Agent 3 can test applications in a browser, revise problems it finds and build agents or scheduled workflows connected to external services. Users can monitor progress and redirect tasks. Extended autonomous runs are an optional beta capability, not a guarantee that every generated application is ready to ship.
- IT之家
Tencent launches CodeBuddy Code CLI and international IDE public testing
CodeBuddy Code brings natural-language development tasks to the terminal, including code generation, refactoring, dependency handling and tests. Tencent also opens international IDE public testing, alongside its existing plugin. The international IDE and CLI share model quotas; trial credits are not a permanent free-service promise.
- IT之家
IndexTTS2 separates vocal identity from emotion for controlled dubbing
IndexTTS2 studies independently controlled voice identity, emotion and duration. Features in a downloadable release should be checked against its documentation, and reference voices require appropriate permission.
- IT之家
InternVL3.5 expands interface, spatial and vector-graphics tasks
InternVL3.5 expands a multi-size visual-language family with interface, spatial and vector-graphics work. Its benchmarks describe specific capabilities; real file and device operations still require authorized, verified workflows.