This week Alibaba has released a series of new AI models, including a unified image generation-and-editing model featured with professional rendering capabilities, an open-sourced embodied foundation model for robotics, as well as an open-sourced lightweight vector database ideal for embedded devices.
Alibaba Releases Enhanced Image Generation Model
Alibaba launched Qwen-Image-2.0, a next-generation foundational image generation model. The model is featured with professional typography rendering, which supports 1k-token instructions for generation of professional infographics such as PPTs, posters and comics; It is also known for stronger semantic adherence, capable of native 2K resolution support for finely detailed realistic scenes, including people, nature, and architecture. The model is also highlighted with improved text rendering, integrated understanding and generation capabilities, unifying image generation and editing in a single mode.
Beyond text-to-image generation, Qwen-Image-2.0 also delivers enhanced image editing capabilities. As a unified generation-and-editing model, its improvements in text rendering and photorealism enhances editing tasks across the board.
Testing on AI Arena shows that Qwen-Image-2.0 achieves superior performance on both text-to-image and image-to-image benchmarks. Global developers can now experience the model on Qwen Chat app.
Alibaba Unveiled Open-sourced Embodied Foundation Model for Robotics
Alibaba DAMO Academy, the research and development institute of Alibaba, unveiled RynnBrain, an open-sourced embodied foundation model based on Qwen3-VL. Moving beyond passive observation, RynnBrain is grounded in the physical world through comprehensive environmental cognition, precise spatiotemporal understanding and task planning. This enables it to perform physics-aware reasoning and execute complex real-world tasks.
RynnBrain is available in dense versions (with 2B and 8B parameters) and a mixture-of-experts (MoE) variant (30B‑A3B), as well as three specialized models: RynnBrain‑Plan for manipulation planning, RynnBrain‑Nav for navigation, and RynnBrain‑CoP for spatial reasoning.
Trained on Alibaba’s latest vision-language model Qwen3‑VL and fine-tuned using DAMO’s proprietary RynnScale system, RynnBrain achieves state-of-the-art (SOTA) results across major benchmarks in embodied cognition, embodied localization, and grounded visual understanding. Its performance is competitive with other leading embodied models such as Gemini Robotics ER 1.5 and Cosmos Reason 2.
RynnBrain is now accessible on Hugging Face, GitHub and the open-source community ModelScope.
Embedded Vector Database for Global Developers
Alibaba recently open-sourced Zvec, an embedded, high-performance, and production-ready vector engine that makes vector retrieval reliable and easy to use. Vector retrieval is critical for AI-native applications, such as Retrieval-Augmented Generation (RAG) for chatbots, recommendation engines and image search.
Designed for end-to-end vector workloads that can run locally with minimal resource usage, Zvec offers rich, high-quality indexing and quantization options to help global developers meet their resource challenges, with deep adaptation across hardware platforms.
While staying easy to use, Zvec provides rich retrieval capabilities, strong resource governance, and excellent search performance. The edge vector database natively supports hybrid search, multi-vector fusion and reranking, enabling robust on-device RAG. Global developers can now access to Zvec via Github and its official page.
