Welcome to the AIbase [AI Daily] column!
Get to know today's major AI events in three minutes a day, helping you understand AI industry trends and innovative AI product applications.
Visit more AI news:https://www.aibase.com/zh
1. Baidu unveils the WENXIN Large Model 4.5 series with ten new models launched!
Baidu officially released the WENXIN Large Model 4.5 series and fully open-sourced it, including ten new models with various parameter configurations. The models were trained and inferred using the PaddlePaddle framework, achieving a FLOPs utilization rate of 47%. They perform excellently in text multimodal benchmark tests and provide a one-stop usage guide and tools, making it easy for developers to fine-tune and deploy. The models have been uploaded to platforms such as Hugging Face and GitHub.
Experience address: https://yiyan.baidu.com
Hugging Face: https://huggingface.co/baidu
GitHub: https://github.com/PaddlePaddle/ERNIE
2. Tongyi Qianwen releases the multimodal unified understanding and generation model Qwen VLo
The Qwen VLo multimodal large model was released, based on the Qwen-VL series upgrade, using a progressive generation method, accurately understanding the world and creating high-quality content, supporting open instruction editing and modification, multi-language instruction capabilities, and can handle image and text input and output. It is currently in the preview stage, with the experience address being the Qwen Chat platform.
Experience address: chat.qwen.ai
3. Alibaba Ovis-U1震撼发布: Multimodal AI Three-in-One, Open Source Empowers Global Developers
The Alibaba International AI team released the Ovis-U1 multimodal large model, with 3 billion parameters, integrating multimodal understanding, text-to-image generation, and image editing functions. It uses an innovative architecture design and is built using Python3.10 and other technology stacks. Compliance checking algorithms were introduced during training, and code model weights are already public, helping to support applications in multiple fields.
Project: (https://huggingface.co/AIDC-AI/Ovis-U1-3B)
4. Huawei opensources PanGu 7B dense and 72B mixture-of-experts models
Huawei open-sources the PanGu 7B dense model, 72B mixture-of-experts model, and Ascend inference technology, practicing the Ascend ecosystem strategy, promoting research on large model technology and industry applications. Related model weights and code have been uploaded to open-source platforms, inviting developers to download, use, and provide feedback.
5. One picture generates a hit video! Meitu MOKI's "AI Creative Advertising" is temporarily free
Meitu MOKI launched the "AI Creative Advertising" feature, allowing users to upload images and select templates to generate professional-level videos. It integrates seven mainstream video generation models. The experience address is www.moki.cn, completing the entire process from creativity to final production.
Experience address: www.moki.cn
6. Gemini 2.5 Pro API returns to free tier, developer community responds enthusiastically
Google's Gemini 2.5 Pro API has returned to the free tier of Google AI Studio. This model has strong multimodal and reasoning capabilities, supports multiple input types. This free return provides developers with innovation opportunities, doubling free computing resources, and the community has responded positively.
7. Douyin's "Deep Research" function is now available on Douyin APP, web version, and PC version for testing
Douyin APP and other platforms have started testing the "Deep Research" function, which can integrate massive deep information to generate research reports or visual web results. Users can get customized reports within minutes by entering instructions and also support一键 converting into podcast formats.
8. Xiaomi's "AI Toolbox" internal testing period ends, service will be suspended starting July 5th
Xiaomi's "AI Toolbox" internal testing has ended, and the service will be suspended starting July 5th. The internal testing collected data feedback, not a discontinued project but rather a strategic planning for data organization. Xiaomi continues to invest in AI exploration and build a multi-layered, full-scenario AI ecosystem.
9. New open-source AI system OmniGen2: Integrates image and text generation like GPT-4o
The Beijing Artificial Intelligence Research Institute launched the open-source system OmniGen2, focusing on text and image generation and editing. It uses an independent decoding path, based on the Qwen2.5-VL-3B transformer, and uses a custom diffusion transformer with a reflection mechanism. Its performance is excellent in multiple benchmark tests and will be released on the Hugging Face platform.
Project: https://huggingface.co/OmniGen2/OmniGen2
10. Zhihu "Direct Answer" upgrades knowledge base function, deeply integrates community content to create an immersive AI Q&A experience
Zhihu "Direct Answer" upgrades the knowledge base function, deeply integrating community content, bringing innovative features such as immersive reading, aiming to provide an immersive AI Q&A experience across multiple scenarios, expanding the influence of answerers' content, and reducing user query costs.
