AI News

At the Volcano Engine Force conference, the Vibe development platform was officially upgraded to "Vibe Coding" and launched for free public testing. This upgrade marks the platform's transformation from an AI application building tool to a platform that empowers developers' creativity, introducing the new Vibe Coding development paradigm. Users only need to describe their business requirements in natural language to automatically generate intelligent agents, workflows, and cross-platform applications.

Apple has announced the open-sourcing of its new AI model, SHARP, which can convert a single 2D photo into a realistic 3D scene in one second. Compared to traditional methods, it boasts a three-order-of-magnitude speedup, achieving a near real-time 3D conversion experience. The model is based on advanced 3D Gaussian splatter technology, using a neural network to predict the position and appearance of millions of 3D Gaussian spheres in a single feedforward.

The MiniMax video team has open-sourced VTP, a key component of its visual generative model, and is the first to propose the scaling property of the tokenizer. This technology significantly improves the performance of downstream generative systems by integrating multiple representation learning methods, linking latent learnability with general representation learning.

OpenAI has released GPT-5.2-Codex, its most advanced agent coding model to date, achieving significant breakthroughs in complex software engineering tasks. The new version significantly improves instruction compliance, long context understanding, and cybersecurity performance, achieving state-of-the-art results in benchmarks such as SWE-Bench Pro and Terminal-Bench 2.0. Its unique context compression technology efficiently handles large codebases, supporting complex refactoring and migration tasks.

Meituan has released and open-sourced its state-of-the-art (SOTA) virtual human video generation model, LongCat-Video-Avatar. The model achieves significant breakthroughs in three key dimensions: motion realism, long video stability, and identity consistency. It natively supports audio-video joint generation and video continuation, and through innovative cross-segment latent space stitching technology, effectively solves the image quality degradation problem in long video generation, stably generating 5-minute high-quality videos.

Aishi Technology and Alibaba Cloud have signed a comprehensive cooperation agreement, under which the two parties will establish deep collaboration in the field of AI video generation. Alibaba Cloud will provide Aishi Technology with full-stack AI support, covering infrastructure and large model services, to facilitate the global deployment of its self-developed video generation model, PixVerse.

Swedish AI programming startup Lovable recently completed a new round of funding, surging its valuation to $6.6 billion, more than tripling in value compared to six months ago, making it one of the most valuable startups in Europe. This round of funding attracted participation from well-known venture capital firms such as Accel, signifying intensified competition in the AI programming tools sector.

Meta has released SAM Audio, the first unified multimodal audio separation model, which supports the accurate extraction of target sounds from complex audio using text, visual, or time-based cues. The model is based on perceptual encoder audiovisual technology, allowing users to click on objects in a video to separate their corresponding audio, or input text commands such as "dog barking" to filter noise.

Step-GUI, a GUI agent model developed by Step-Star, has been fully upgraded. The new model supports over 200 task scenarios, features longer inference steps, stronger semantic understanding, and improved generalization capabilities. It can be used on multiple platforms including mobile phones, PCs, and automobiles, and can be deployed in as little as 10 minutes. The company has released its API for free and published a technical report and the GUI-MCP protocol.

ByteDance officially released Seed 1.8, a general-purpose agent model. The model integrates diverse capabilities such as search, code generation, and graphical user interface interaction, supporting both text and image input and possessing native visual understanding functionality. It has demonstrated excellent performance in multiple benchmark tests, achieving a score of 61.9 on the OSWorld GUI Agent task and 67.6 on the BrowseComp-en search task, both placing it among the industry leaders.

Google officially released the high-speed, low-cost Gemini 3 Flash model. The model has demonstrated outstanding performance in multiple benchmark tests, surpassing its predecessor, Gemini 2.5 Pro, and approaching cutting-edge models such as Gemini 3 Pro and GPT-5.2 in some metrics. The new model features state-of-the-art inference and multimodal capabilities, with a 3x faster response time, while significantly reducing costs.

Tencent officially released WorldPlay 1.5, the first world model in China to support real-time interactive experiences. Users can generate their own 3D interactive world by inputting text or images, and freely explore it using a keyboard, mouse, or gamepad. The model features 24FPS real-time generation, minute-level 3D geometric consistency, and spatial memory capabilities, maintaining consistency upon returning to the scene after leaving.

Xiaomi officially released the open-source large-scale model MiMo-V2-Flash. The model adopts the MoE architecture, with a total of 309 billion parameters and 15 billion active parameters. It performed exceptionally well in multiple benchmark tests, with its programming capabilities being particularly outstanding. In the SWE-bench test, it achieved a resolution rate of 73.4%, surpassing all other open-source models. Through innovative technologies such as hybrid sliding window attention and multi-token prediction, the model achieves a high inference speed of 150 tokens/second and extremely low inference cost.

OpenAI has released its new flagship image generation model, GPT Image 1.5. The model has made significant breakthroughs in cue word adherence, detail preservation, and text rendering, and can accurately generate or edit images while maintaining the original image's core features such as lighting and composition.

ByteDance has released Seedance 1.5 pro, a new generation audio and video creation model that enables joint audio and video generation. The model supports the synthesis of audio and video from text and image-guided generation, achieving breakthroughs in accurate audio-visual synchronization, multilingual and dialectal lip-syncing, cinematic camera control, and dynamic narrative performance.

Jimeng AI Web Version has been fully upgraded, creating a one-stop "AI Film Studio". Key updates include: the launch of the "Video 3.5 Pro" model based on Seedance 1.5 Pro, supporting joint audio and video generation and dialect lip-syncing; Intelligent Multi-Frame 2.0 enabling "video + video" stitching and fine-grained editing of clips, supporting a single take of up to 20 clips; and Image 4.1 and 4.5 models respectively enhancing graphic design capabilities and portrait aesthetics.

SenseTime released its AI office assistant, "Little Raccoon 3.0," achieving three major leaps from draft generation to final product delivery. The assistant boasts the ability to generate high-quality PPTs with a single click, supporting full-element editing of charts, images, and more, delivering ready-to-use presentations. Its "long-chain thinking" capability supports the processing of millions of data points in seconds and multimodal analysis, enabling deep understanding of complex tasks.

Alibaba Cloud's Tongyi Bailing open-source speech twin models Fun-CosyVoice3 and Fun-ASR. Fun-CosyVoice3 supports voice cloning and emotion control across 9 languages and 18 dialects with just 3 seconds of audio, reducing first-packet latency by 50% and the error rate for mixed Chinese and English speech by 56.4%. Fun-ASR achieves 93% accuracy in noisy environments, supports mixed speech in 31 languages and dialect recognition, and reduces the first-word latency of streaming recognition to 160 milliseconds.

NVIDIA has released the Nemotron 3 open-source AI model family, including Nano, Super, and Ultra versions. The Nano version (30 billion parameters) is already available, employing a hybrid expert architecture and claiming four times the throughput of its predecessor, a 60% reduction in inference token generation, and a 1 million token context window. This family is designed for multi-agent systems, aiming to address challenges such as communication overhead and inference costs.

Alibaba has released Tongyi Wanxiang 2.6 multimodal model, the first video generation system in China to support role-playing functionality. The model supports synchronized audio and video, multi-camera storytelling, and audio-driven generation, capable of producing 15-second high-definition videos directly. The role-playing function allows users to upload reference videos; the AI can replicate the character's appearance and voice to generate single or multi-person performance content. In terms of image generation, the model improves its ability to blend artistic styles and depict details, supporting commercial-grade creation such as mixed text and image layouts and multi-image fusion.

LiblibAI platform globally launches the Tongyi Wanxiang Wan 2.6 video generation model, supporting three core functions: video reference generation, audio-visual synchronization, and intelligent multi-camera scheduling. The model can directly output 15-second 1080P high-definition videos, achieving complete narrative generation without secondary editing. Its video reference generation capability can replicate the appearance and voice of any character within 5 seconds, supporting audio-visual synchronization in single/dual performances and multi-person dialogue scenes.

Google has released upgraded versions of its Gemini TTS model, Gemini 2.5 Flash and Pro text-to-speech models, improving expressiveness, speech rate control, and performance in multi-speaker scenarios. The new model offers better control over style, intonation, and pronunciation, adjusting speech rate according to context—speeding up when excited and slowing down for emphasis—and maintaining consistent character voices in multilingual dialogues for a more realistic experience.

Ant Group's AI health assistant, AQ, has been updated and renamed "Ant Afu" in Chinese, while also launching the "Health Goals" feature. The app offers health consultations, image interpretation (reports, medical records, prescriptions, and medicine boxes are all supported), personal and family health record management, as well as medical and health services such as appointment booking and cloud-based medical companionship.

SenseTime released Seko 2.0, an intelligent agent for generating multi-episode series. As the industry's first AI tool integrating creation and editing, it is designed specifically for short-episode creators. This version solves the problem of character consistency across multiple series using SekoIDX technology. SekoTalk achieves the industry's first function for accurate lip-syncing between two or more people, reducing the traditional 3-month production cycle for 50 episodes by 80%-90%. Simultaneously, it open-sourced the LightX2V inference framework, enabling the generation of 5-second videos in just 5 seconds on consumer-grade graphics cards, achieving compatibility with domestic chips.