AI News

ByteDance has released its next-generation multimodal AI model, Seedream 5.0 Lite. This version doesn't prioritize higher resolution, but rather significantly enhances the model's "thinking" and "reasoning" capabilities, including visual reasoning, information visualization, style transfer, and precise image editing. A key highlight is the introduction of real-time retrieval enhancement capabilities, enabling content creation based on real-time online information.

The AgentScope team officially released the personal AI agent "CoPaw," positioned as a "collaborative personal AI agent workbench," becoming a dedicated intelligent partner that users can deploy locally. Compared to its predecessor, OpenClaw, CoPaw focuses on simplified installation and a unified experience, supporting both quick deployment on users' local computers (macOS/Windows/Linux) via simple commands and one-click cloud deployment on "Magic Space."

OpenAI, in collaboration with Cerebras, has launched a new model, GPT-5.3-Codex-Spark, designed for real-time programming. Optimized for high-speed software development scenarios, the model can perform inference at a rate exceeding 1000 tokens per second, achieving near-instantaneous coding feedback. This addresses the issues of excessive waiting time and reduced control experienced by developers during long periods of AI-driven autonomous programming.

Google has released a major upgrade to its Gemini 3 Deep Think model. The new model has broken records in multiple top benchmark tests, notably achieving an Elo rating of 3455 on the competitive programming platform Codeforces, ranking eighth globally, with only seven top programmers able to surpass it in competitions. The model has also achieved gold medal-level performance in several academic competitions, including the ARC-AGI-2 inference benchmark, "Humanity's Last Test" (without tools), and the written exams for the Physics and Chemistry Olympiads.

AI company Anthropic (the company that developed Claude) announced the completion of a massive $30 billion Series G funding round, bringing its post-money valuation to $380 billion, setting a new record for a single-round equity financing for an AI company. This round was led by Singapore's sovereign wealth fund GIC and Coatue, with participation from several other top-tier investment firms.

MiniMax officially released the M2.5 model, whose performance in high-level productivity scenarios such as programming, searching tools, and office work reaches or surpasses the industry's state-of-the-art (SOTA) level. The model significantly improves task decomposition capabilities and inference efficiency, completing tasks on the programming benchmark SWE-Bench Verified 37% faster than the previous generation M2.1.

Meituan's LongCat has officially launched its "Deep Research" intelligent agent feature, leveraging real transaction data and industry insights from Meituan's local life services platform to generate personalized dining, entertainment, and travel guides for users. The intelligent agent boasts long-term task planning and multi-tool invocation capabilities, supports 400 rounds of interaction and 256K context processing, and has achieved state-of-the-art (SOTA) performance in authoritative benchmarks such as BrowseComp and MCP-Atlas.

ZCode, a new version of Zhipu AI, is the first to define an ADE (Agent Development Environment), transforming complex CLI interactions into a multi-agent collaboration platform through a graphical interface. The environment supports visual configuration management, cross-device remote development, and multi-task parallel processing. It features a built-in real-time preview browser and session-level version management, and can flexibly utilize mainstream toolchains such as Claude Code.

Zhipu AI has open-sourced its GLM-5 large-scale model, expanding its parameter size to 744B (40B activations) and pre-training data to 28.5TB. It achieves state-of-the-art (SOTA) performance in both coding and agent capabilities, with overall performance approaching that of Claude Opus 4.5. This model leads open-source models in programming benchmarks such as SWE-bench and Terminal Bench, capable of end-to-end completion of complex systems engineering and long-range agent tasks, supporting resource planning scenarios such as vending machine operation simulation.

iFlytek has released the Spark X2 large-scale model, trained using domestically produced computing power. Its general capabilities have been comprehensively upgraded, benchmarking against top international models in core metrics such as mathematics, reasoning, and language understanding, while maintaining a leading position with over 130 languages supported. The Spark X2 employs a 293B MoE sparse architecture and multiple engineering optimizations, achieving a 50% improvement in inference performance compared to the X1.5, and can be deployed on a single domestically produced Ascend server.

Alibaba DAMO Academy has open-sourced RynnBrain, the industry's first embodied brain foundation model, which possesses spatiotemporal memory and physical spatial reasoning capabilities. The model series ranges in parameter size from 2B to 30B, employing the MoE architecture for efficient reasoning; activating only 3B parameters surpasses the performance of a 72B model. It comprehensively outperforms top models from NVIDIA, Google, and other leading companies in 20 embodied benchmark tests, overcoming the challenge of robots' "forgetting what they see" in physical world cognition.

The domestically developed AI application Loopit has sparked heated discussions on overseas social media. Developed by Chen Weipeng's team, co-founder of Baichuan Intelligence, the product focuses on "AI interactive content generation." Users can automatically generate clickable, swipeable, and interactive content based on text descriptions, and can also "modify" and recreate other people's works.

DaXiao Robotics recently completed its angel round of financing, led by Ant Group, with participation from Qiming Venture Partners, Jinjing Capital, and other institutions. The funds will be used to advance its "human-centered" ACE embodied full-stack R&D paradigm, accelerate environmental data collection, and develop the Enlightened World Model 3.0.

Tencent Hunyuan has launched the industry-grade 2-bit edge model HY-1.8B-2Bit, with an equivalent parameter count of only 0.3B and a memory footprint of 600MB, smaller than that of common mobile applications. The model is trained using 2-bit quantization, maintaining the original model's full thinking capabilities while increasing generation speed by 2-3 times, allowing for seamless deployment on edge devices such as mobile phones and smart home devices.

Alibaba has released Qwen-Image-2.0, a new generation image generation and editing model that supports 1K token-long text input and 2K high-resolution output. This model unifies image generation and editing functions for the first time, ranking third globally in image generation with a score of 1029 in AI Arena's evaluation, and achieving a score of 1034 in image editing. Its key strengths lie in its excellent Chinese text rendering capabilities and understanding of complex instructions, enabling the generation of high-quality images such as professional PPT presentations and multi-panel comics.

ByteDance's image generation model Seedream 5.0 has officially launched and is available for a limited-time free trial on platforms such as CapCut, Xiaoyunque, and others. The new model supports 2K/4K resolution output, significantly improves the accuracy of prompt word understanding, detailed texture rendering, and intelligent reasoning capabilities, and adds online search functionality.

Tencent has launched WorkBuddy, a full-scenario AI office assistant, and officially opened applications for internal testing. Positioned as a smart desktop workbench, the product supports natural language interaction and can autonomously manipulate local files after understanding commands, performing complex tasks such as PPT generation, data organization, and knowledge base construction. Unlike chatbots, WorkBuddy emphasizes "execution capability," delivering results like a real colleague.

The Zhipu Qingyan APP has officially launched the "Learning Partner" feature, allowing users to learn anytime, anywhere, utilizing fragmented time. Within two weeks of its web launch, over 12,000 learning projects have been created. Now fully available on mobile, no invitation code is required to experience it.

Skywork has launched SkyBot, touted as the world's most user-friendly OpenClaw tool for beginners. This product is a cloud-based transformation of the popular AI agent OpenClaw, allowing users to access a 24/7 cloud-based AI assistant without needing to configure hardware or command lines.

Beijing Shengshu Technology announced the completion of its Series A+ financing round, raising over 600 million yuan, setting a new record for single-round financing in the domestic video generation field. Founded by a team with Tsinghua University backgrounds, the company's Vidu Q3 video generation model has demonstrated outstanding performance in professional film and television production scenarios, ranking second globally. Its clients include companies such as Sony, Tencent, and ByteDance.

AI marketing platform PallasAI has completed a multi-million yuan funding round, exclusively invested by Youzan. The company launched the industry's first standardized GEO AI Agent product, which helps SMEs quantify their brand exposure within the AI ecosystem through three main functions: visibility scoring, marketing graph, and Agent content collaboration.

Anthropic released its next-generation AI model, Claude Opus 4.6, featuring a 1 million-token context window and adaptive thinking capabilities. It has demonstrated exceptional performance in programming, financial analysis, and office software operation: its 16 AI agents wrote a C compiler capable of compiling Linux within two weeks, independently discovered over 500 zero-day vulnerabilities, and surpassed GPT-5.2 in the GDPval-AA benchmark.

OpenAI has launched its next-generation programming model, GPT-5.3-Codex, claiming it to be the world's most powerful intelligent agent programming model. The model achieved state-of-the-art (SOTA) scores in SWE-Bench Pro and Terminal-Bench 2.0 tests, with a programming score 11.9% higher than Claude Opus 4.6. It possesses debugging, deployment, and office software operation capabilities, with a 25% speed improvement, and can participate in its own development and optimization.

Mistral AI's next-generation speech-to-text model, Voxtral Transcribe 2, comes in two versions. Voxtral Mini supports high-precision transcription in 13 languages and features speaker separation, word-level timestamps, and context bias, making it suitable for batch transcription. Voxtral Realtime is designed for real-time scenarios with latency as low as 200 milliseconds, making it ideal for interactive applications such as voice assistants.