AI News

Ant Financial's large model team has released Ling-2.6-flash, with a total of 104B parameters and 7.4B activation parameters. It adopts a hybrid attention architecture of MLA+Lightning Linear and sparse MoE. The model achieves an inference speed of 340 tokens/s in a 4-card H2O environment, with token consumption only about 1/10 of its competitors. It achieves state-of-the-art performance on Agent benchmarks such as BFCL-V4 and SWE-bench Verified.

SpaceX, Elon Musk's commercial spaceflight giant, has announced a deep partnership with AI programming unicorn Cursor, with plans to acquire it for $60 billion later this year. If the acquisition ultimately falls through, SpaceX will pay an additional $10 billion for the partnership. The two companies will combine Cursor's leading products with SpaceX's Colossus supercomputer, which has the equivalent computing power of a million H100 supercomputer, to jointly build the world's most powerful programming AI.

Loopit, an AI-powered interactive content community, has completed a new round of financing of $50 million, led by Garena, a leading global mobile game developer, with participation from BlueRun Ventures and other institutions. This brings its total financing this year to nearly $100 million. Founded by Chen Weipeng, former co-founder of BaiChuan Intelligent, Loopit reached number 8 on the Google Play global charts and number 1 on the entertainment charts within two months of its launch, earning praise from Elon Musk.

OpenAI officially launched ChatGPT Images 2.0 (GPT-Image-2), its first image generation model with thinking capabilities. The model excels in accuracy, timeliness, consistency, and visual coherence, generating highly realistic app screenshots, TikTok video interfaces, product advertisements, and academic paper posters using only simple prompts. Its Chinese text rendering capabilities have also been significantly improved.

Alibaba's Tongyi Labs announced an official upgrade to AgentScope Tuner, a one-stop automated optimization engine designed specifically for Agentic AI. The product offers three main capabilities: Prompt tuning, model selection, and enhancement fine-tuning, covering the entire lifecycle from initial lightweight optimization to later-stage deep tuning. Key highlights include a native Agent closed loop, a unified API design paradigm, and zero code modification costs.

Tencent's QClaw overseas version has officially launched its internal beta test, and is now available in multiple countries and regions including the United States, Canada, and Singapore. The product emphasizes zero barriers to entry, no deployment required, and immediate usability after download. It supports remote control of computers via WhatsApp or Telegram to complete tasks such as tax filing, fitness planning, and social media management. 99% of the code for the overseas version of QClaw was written in-house by QClaw in just 5 days.

The Dark Side of the Moon open-sources the Kimi K2.6 model, boasting industry-leading code, long-term task execution, and agent clustering capabilities. The model performs on par with or better than closed-source models such as GPT-5.4 and Claude Opus 4.6 in multiple benchmark tests, including Humanity's Last Exam and SWE-Bench Pro. K2.6 can be continuously coded for 13 hours with over 4000 lines of code modifications. The agent cluster supports 300 sub-agents completing 4000 collaborative steps in parallel, and can support continuous autonomous operation for up to 5 days.

Alibaba has launched Fun-ASR 1.5, an end-to-end speech recognition model that covers 30 languages and supports automatic language switching and cross-language mixed speech. The model covers the seven major Chinese dialect systems, with a 56.2% reduction in word error rate compared to the previous version. The newly added classical poetry recognition feature achieves an accuracy rate of 97%. Fun-ASR 1.5 uses the MoE architecture and supports intelligent punctuation prediction and text normalization for numbers, dates, etc. It is now available on Alibaba Cloud's Bailian and Moda communities.

Alibaba's Tongyi Qianwen has released an early preview version of its next-generation flagship model, Qwen3.6-Max-Preview. Compared to its predecessor, the model shows significant improvements in agent programming, world knowledge, and instruction compliance, achieving top scores in six programming benchmarks. Users can experience it online through Qwen Studio or via the Alibaba Cloud Bailian API, which is compatible with the OpenAI and Anthropic protocols.

Anthropic launched Claude Design, an AI-native visual design platform powered by Claude Opus 4.7, directly challenging traditional design software like Adobe and Figma. The product boasts three core capabilities: automatic extraction of brand design systems, support for multimodal input including documents, screenshots, and code libraries, and one-click generation of runnable code from design drafts. Following the announcement, the stock prices of design software giants like Adobe, Figma, and Wix collectively plummeted.

The Alibaba Tongyi Qianwen team has open-sourced the Qwen3.6-35B-A3B sparse MoE model, with a total of 35 billion parameters but only 3 billion activations. Optimized specifically for agent programming, its performance significantly surpasses its predecessor and rivals that of larger, denser models. The model supports multimodal reasoning and both thinking and non-thinking modes, demonstrating outstanding performance in benchmark tests such as code generation and tool invocation.

Anthropic officially launched Claude Opus 4.7, positioning it as the most powerful and widely usable large-scale model currently available. Core upgrades focus on complex task execution, high-definition visual understanding, and stability in long-chain workflows. Visual capabilities have achieved a qualitative leap, jumping from 54.5% to nearly perfect 98.5% in the XBOW test; programming capabilities reach 80.5% in the SWE-bench; and the BFS test for millions of tokens with long context has improved from 41.2% to 58.6%.

AutoClaw, a subsidiary of Zhipu AI, has officially launched its self-evolution mechanism and Skill Store. The self-evolution function automatically identifies user corrections, preferences, and lessons learned from failures, which are then permanently stored after approval, allowing the Agent to become increasingly adept at understanding users. The platform also simultaneously launched the GLM Office Skills suite, based on GLM-5.1, supporting various design scenarios such as PPT and Word, intelligent self-checking, and format conversion, enabling one-click generation of accompanying office materials.

NetEase Youdao officially launched its first AI knowledge base product, "Youdao Treasure Trove." Positioned as an "external super knowledge base," the product supports document storage, intelligent question answering, and one-click generation of PPTs, podcast scripts, research reports, and other content. Core scenarios include: mastering knowledge in an unfamiliar field in 30 minutes and generating a 20-page PPT; automatically organizing project documents for the entire year and generating an annual summary in 5 minutes; uploading celebrity biographies to simulate a "celebrity board meeting," providing decision-making suggestions, and generating podcasts.

HiDream.ai, a multimodal generative AI company, announced the completion of a new round of financing exceeding 500 million yuan. The funds will be used for the development of native multimodal world models, enterprise intelligent agent products, and global market expansion. The company's open-source model, HiDream-I1, topped the Artificial Analysis leaderboard, with over 2 million downloads. Currently, its products cover 30 million professional users and over 40,000 enterprise clients, forming a closed loop across three major scenarios: commercial marketing, film and television creation, and social media content.

StepAudio 2.5 TTS, a next-generation speech generation model, has been officially launched by StepAudio, featuring three core capabilities: global context control, text-based context control, and zero-shot reproduction. Users can precisely control details such as tone, rhythm, pauses, and emphasis in speech through natural language processing, achieving a leap from "reproducing voices" to "creating expressions." The model supports zero-shot reproduction of any timbre, generating high-quality speech without retraining.

Alibaba's ATH Innovation Division has launched HappyOyster, an open-world modeling product, which is now officially in closed beta testing. The product supports real-time building and interaction, offering two gameplay modes: "Directing" (real-time director) and "Wandering" (world roaming). Users can create within an infinitely generated video stream or freely explore an extended space. Applications are now open on the official website; users can join the waitlist to obtain early access to the closed beta.

Tencent's HY-World 2.0 3D world model has been officially released and open-sourced. The model supports multimodal input including text, images, and videos, and can generate, reconstruct, and simulate a complete 3D world with a single click. It supports exporting editable assets such as Mesh/3DGS/point clouds and seamlessly integrates with Unity/UE game engines. Compared to version 1.0, 2.0 adopts the WorldMirror 2.0 architecture, significantly improving image detail and realism, and supporting free exploration and physical collision detection in character mode.

MiniMax officially launched MaxHermes, the world's first cloud-based sandbox for Hermes, a cloud-based self-evolving AI assistant built on the Hermes Agent. The core innovation of the product is its closed-loop learning mechanism: after completing a complex task, it automatically extracts reusable skills for continuous self-iteration, equipped with persistent memory and multi-sub-agent parallel capabilities. The product offers zero-barrier deployment and is already integrated with IM channels such as Lark, DingTalk, and WeChat Work, supporting Token Plan deduction for consumption. (MiniMax Technology)

Mianbi Intelligence has launched Lantay, a professional-grade document creation platform, which is now in public beta. Drawing inspiration from Cursor's design philosophy, the product utilizes a Vibedocing interaction paradigm to enable collaborative human-computer creation, employing a three-in-one layout encompassing a navigation area, workspace, and dialogue area. The platform supports 16 file formats, can batch-integrate up to 200 documents at a time, and has a single file size limit of 100MB. Lantay is primarily targeted at heavy document users in highly rigorous industries such as law, finance, and publishing.

Meta announced a five-year deal with chip giant Broadcom (until 2029), with an initial investment of over 1GW of computing power for custom AI chips. Zuckerberg declared the goal is to "provide personal superintelligence for billions of people." Meta plans to launch four generations of its self-developed MTIA chips (300/400/450/500) within two years, establish a superintelligence lab, and appoint Alexandr Wang, founder of Scale AI, as its chief AI officer.

Alibaba's AITH Business Group has launched its first AI development tool, Meoo. The product integrates four top-tier models: Qianwen, Kimi, GLM, and MiniMax, and includes core services such as Alibaba Cloud database and storage. Users need no programming background; they can describe their requirements using natural language, and the tool can automatically generate complete front-end and back-end code in as little as one minute, supporting one-click deployment on Alibaba Cloud.

World Labs, under Fei-Fei Li's company, has open-sourced Spark 2.0, a 3D Gaussian sputtering rendering engine that supports streaming over 100 million splats of 3D worlds to desktops, mobile phones, and VR devices via WebGL2. Built on Three.js, the engine employs continuous LoD Gaussian sputtering trees, progressive streaming, and virtual memory technologies to achieve real-time rendering of ultra-large-scale scenes on web pages, addressing the pain points of traditional engines such as supporting only single objects, lack of dynamic animation, and poor device compatibility.

Baidu's Wenxin platform has officially open-sourced its text rendering model, ERNIE-Image. With only 8 parameters, it achieves state-of-the-art (SOTA) performance in open-source environments, rivaling commercial closed-source models like Nano Banana in text rendering and complex instruction compliance. The model requires only 24GB of GPU memory to run, supports accurate glyph generation for multiple languages including Chinese, English, Japanese, and Korean, and is available on ComfyUI with a GGUF quantization solution. Related weights and inference code have been open-sourced on Hugging Face.