AI News

The open-source long-term memory operating system EverMemOS, developed by the Shanda team led by Chen Tianqiao, is designed specifically for AI agents and addresses the memory fragmentation problem caused by fixed context windows in large language models. Based on the human brain's memory mechanism, the system employs a four-layer architecture (agent layer, memory layer, index layer, and interface layer), achieving high scores of 92.3% and 82% in LoCoMo and LongMemEval-S benchmarks, respectively, surpassing industry benchmarks.

Alibaba officially launched the Qianwen app, built on Qwen's most powerful model and fully benchmarked against ChatGPT. Leveraging the Qwen open-source ecosystem, the app has been downloaded over 600 million times globally, supports 119 languages, and possesses capabilities such as intelligent multi-application scheduling and multimodal interaction, becoming an AI productivity tool for users worldwide.

US-based generative AI startup Gamma has raised $68 million in Series B funding, valuing the company at $2.1 billion. Founded in 2020, Gamma is a generative AI design platform that enables the rapid creation of presentations, documents, and social media posts, supporting 22 AI image models and more than 60 languages.

OpenAI announced a pilot program for its ChatGPT group chat feature in select regions, including South Korea and New Zealand. Based on the GPT-5.1 model, the feature allows users to collaborate with ChatGPT members, inviting friends, family, or colleagues to a shared space for planning, decision-making discussions, or brainstorming.

AI programming tool Cursor (parent company Anysphere) has completed a $2.3 billion Series D funding round, bringing its post-money valuation to $29.3 billion, making it the world's first AI programming startup with a valuation exceeding 200 billion yuan. This round was led by Accel and Coatue, with participation from NVIDIA, Google, and others. In the past year, Anysphere's valuation has increased more than tenfold, and its annualized revenue has surpassed $1 billion.

Dexmal, a company specializing in embodied intelligence, announced the completion of its Series A+ funding round, raising hundreds of millions of yuan, with Alibaba as the sole investor. Previously, its Series A funding round was led by NIO Capital, bringing the total funding raised in these two rounds to nearly 1 billion yuan. The funds will be used for the research and development and implementation of intelligent robot software and hardware technologies. Dexmal focuses on the research and development of embodied intelligence technology, and its self-developed multimodal embodied intelligence model MMLA can achieve cross-scenario intelligent generalization.

LMArena, a leading global AI evaluation platform, updated its Code Arena (programming and code generation) rankings, with Zhipu GLM-4.6 ranking first globally alongside top models such as Anthropic Claude and OpenAI GPT-5.

World Labs, founded by "AI Godmother" Fei-Fei Li, has officially launched its first commercial multimodal world model product, Marble. Users can generate high-fidelity, editable 3D virtual worlds using text, images, videos, or 3D layouts, and export them to Gaussian sputtering, mesh, or video formats.

ByteDance's TRAE team officially released the full version of TRAE SOLO. Positioned as a "responsive programming intelligent agent," this version supports complex project development and adds built-in intelligent agents SOLO CODER and SOLO BUILDER. It features context compression, multi-task parallelism, and visual tool invocation, efficiently handling project setup from scratch and iterative optimization from 1 to 100. It is now available to global users for a limited time with a free trial.

Baidu officially released Wenxin Big Model 5.0 at its World Conference. The model boasts 2.4 trillion parameters, employs native full-modal unified modeling technology, and supports multimodal input and output including text, images, audio, and video. It excels in multimodal understanding, creative writing, and instruction compliance, achieving a globally leading overall capability.

OpenAI's upgraded AI version, GPT-5.1, focuses on optimizing the dialogue experience rather than simply improving performance metrics. The new version includes two main models: Instant and Thinking. The former emphasizes more human-like communication (such as adaptive thinking and eight style presets), while the latter enhances professional reasoning capabilities (simplifying terminology). Compared to its predecessor, GPT-5.1 shows significant improvements in emotional interaction and command understanding.

AI design platform Lovert has officially launched the "Edit Elements" feature. After users upload their finished posters, the AI can automatically break them down into independent editable layers such as text, main content, and background, supporting text modification, element replacement, and real-time preview.

Chat2DB founder Ji Pengfei and his team officially launched ZOER.AI, a full-stack application building platform that focuses on "Database-First". It aims to solve the shortcomings of traditional AI programming tools in backend capabilities, and achieve full-process automation from the data layer to the underlying services by intelligently designing database structures and automatically generating secure backend logic and frontend interfaces.

Excellent Vision has completed a RMB 100 million Series A1 financing round, jointly invested by Huawei Hubble and Huakong Fund. This is its third round of financing within two months. Excellent Vision's technology solves problems such as data scarcity and simulation errors in the field of physical AI. Its CEO predicts that the "Physical World ChatGPT moment" will arrive within 2-3 years.

OiiOii is a revolutionary animation creation tool. As the world's first animation creation agent system, it has built-in 7 intelligent agents, including art director, screenwriter, and storyboard artist. Users only need to input their ideas or upload reference images to generate scripts, storyboards, character designs, and complete animated short films with one click. It supports 161 styles, covering various scenarios such as short films, music videos, and comics to videos, greatly reducing the threshold for animation creation. It is currently in the internal testing phase and is free to use for a limited time.

The KaLM-Embedding-Gemma3-12B-2511 model, developed by Tencent's WeChat team, ranked first globally in the authoritative MTEB multilingual general-purpose embedding model evaluation. With 12 billion parameters, supporting multiple vector dimensions from 3840 to 64, the model possesses outstanding cross-lingual semantic understanding and retrieval capabilities. It excels in multilingual semantic alignment, data quality optimization, and innovative training strategies, providing robust semantic foundation support for multilingual application scenarios.

Baidu's Xiaodu AI Glasses Pro has officially gone on sale, priced at 2299 yuan. It is the second major internet company in China to officially launch AI glasses after Alibaba's Quark AI Glasses S1. Xiaodu AI Glasses Pro is an AI photography glasses, similar in type to Xiaomi AI Glasses. It is not the more "complete" version of AI+AR glasses currently available. It integrates a multimodal AI model, and its functions include taking photos, identifying songs, intelligently matching playlists, AI translation, AI object recognition, AI memos, and AI recording.

Meta AI has launched Omnilingual ASR, the world's first automatic speech recognition system to support over 1,600 languages, with 500 languages receiving AI speech recognition support for the first time. Through large-scale pre-trained models and contextual learning techniques, the system can expand to new languages with only a small number of audio-text samples, achieving a character error rate of less than 10% for 78% of the languages.

SenseTime has achieved a significant breakthrough in the field of spatial intelligence, officially releasing and open-sourcing its SenseNova-SI series models. In multiple authoritative evaluations of spatial understanding and reasoning tasks, SenseNova-SI not only significantly outperforms similarly sized open-source multimodal models but also surpasses the performance of top international closed-source models such as GPT-5 and Gemini 2.5 Pro.

Moonshot AI has released Kosong, a new open-source AI agent development framework. With the core philosophy of "simplifying complexity and unleashing creativity," it provides lightweight and highly scalable underlying support for next-generation intelligent agent applications. The framework effectively addresses development pain points such as tool fragmentation and interface incompatibility by encapsulating standardized components through a unified LLM abstraction layer and incorporating an asynchronous tool orchestration engine.

ByteDance has officially launched Doubao-Seed-Code, optimized for complex programming tasks. The model supports a 256k long context, is compatible with the Anthropic API and mainstream IDEs, and its performance is second only to Claude Sonnet 4.5. The overall cost of use is reduced by 62.7%, and it has topped the SWE-Bench Verified leaderboard.

OpenAI has released GPT-5-Codex-Mini, a "cost-effective" version of GPT-5-Codex. Developers receive approximately four times the usage credit. In the SWE-bench Verified test, GPT-5-Codex-Mini scored 71.3%. This version is suitable for lightweight engineering tasks or situations nearing the speed limit.

Baidu Wenxin 5.0 Preview ranked second globally and first in China in the LMARaena global text writing arena with a score of 1432, comparable to top models such as OpenAI GPT-4.5 and Anthropic Claude. Real-world testing shows its outstanding performance in creative writing, long text comprehension, and complex instruction adherence, generating content that is both poetic and logically rigorous.

Duiyou AI Reactor has officially launched its AI-powered all-in-one creative canvas, which is available for free to all Duiyou users. Duiyou AI integrates functions such as text-to-image generation, image-to-image generation, intelligent filters, vector editing, font design, high-definition zoom, intelligent background removal, and partial redrawing, providing a one-stop AI design solution that automates the entire process from inspiration to delivery.