AI News

OpenAI has launched a new "Shopping Research" feature for ChatGPT, replacing the traditional consumer decision-making process of e-commerce platforms and content communities. The feature is optimized based on the GPT-5 mini model, using interactive questioning to understand user preferences, combining memory functionality to generate personalized buying guides, and directly providing product links and visual comparison interfaces.

AMD, in collaboration with IBM and AI startup Zyphra, has successfully trained its first large-scale hybrid expert (MoE) model, ZAYA1, using AMD hardware after more than a year of development. The model is entirely built on the AMD hardware ecosystem, with training conducted on the IBM Cloud using technologies such as the AMD Instinct MI300X GPU. The training cluster consists of 128 nodes, totaling 1024 GPUs, achieving a performance exceeding 750 PFLOPs.

Tencent's Hunyuan Open Source Lightweight OCR Model, HunyuanOCR, boasts only 1B parameters yet achieves state-of-the-art (SOTA) performance in several core capabilities. Based on Hunyuan's native multimodal architecture and employing an end-to-end design, the model achieved a score of 94.1 on the complex document parsing benchmark OmniDocBench, surpassing mainstream models such as Gemini 3 Pro.

Anthropic released its new generation of large-scale model, Claude Opus 4.5, achieving a major breakthrough in programming capabilities. It became the first model to score over 80% in the SWE-Bench Verified real-world software engineering test and ranked first in seven out of eight programming languages. Core upgrades include a precise understanding of ambiguous requirements, autonomous bug localization capabilities, and an innovative "effort" parameter control system that significantly reduces token consumption while maintaining performance.

The Olmo 3 series of open-source large-scale language models launched by the Allen Institute for Artificial Intelligence (AI2) includes Olmo 3-Base (7B and 32B parameters), Olmo 3-Think, Olmo 3-Instruct, and Olmo 3-RL Zero, which are respectively good at tasks such as programming, reading comprehension, mathematical problem-solving, complex reasoning, multi-turn dialogue, and reinforcement learning.

Alibaba's AI assistant "Qianwen" has officially launched its new domain name, qianwen.com, further enhancing user access convenience. Users can now experience Qianwen directly through the web browser, achieving a seamless user experience consistent with the app. The Qianwen web browser also offers more Qwen3 series model options for professional users.

Meta introduces the WorldGen system, which generates interactive and navigable 3D worlds with just a single text prompt. The system integrates procedural logic reasoning and diffusion models to output geometrically rigorous and visually rich 3D scenes suitable for games, simulations, and more. It can generate 50m x 50m fully textured scenes with a high degree of consistency between style and structure.

Singapore-based AI fintech company RockFlow has completed a multi-million dollar funding round, led by Ant Group. The company launched "Bobby," the world's first AI agent for financial transactions, which can understand users' investment intentions through natural language and autonomously complete the entire process from data analysis to real-time trading. This funding round will accelerate the innovative application of AI technology in financial services, driving the development of complex financial services towards greater accessibility and personalization.

Xiaomi has released MiMo-Embodied, the industry's first cross-domain model that integrates autonomous driving and embodied intelligence, and has fully open-sourced it. The model unifies the three core tasks of embodied intelligence with the three key tasks of autonomous driving, enabling bidirectional knowledge transfer between indoor interaction and road decision-making capabilities. It has demonstrated outstanding performance in 29 core benchmark tests, comprehensively surpassing existing open-source and dedicated models.

Tencent's HunyuanVideo 1.5 open-source video generation model, with only 8.3B parameters, can generate 5-10 second high-definition videos. Based on the DiT architecture, the model supports text-based and image-based video generation in both Chinese and English. It possesses strong command understanding capabilities and can precisely control camera movement, facial expressions, and physical motion. It natively supports 480p/720p output, which can be upscaled to 1080p via super-resolution. It runs smoothly on consumer-grade graphics cards with 14GB of VRAM, significantly lowering the barrier to entry for users.

Google has released Nano Banana Pro, a professional-grade image generation model deeply integrated with Gemini 3's multimodal capabilities. The model supports 2K/4K high-resolution output, can fuse up to 14 reference images while maintaining consistency across 5 individuals, and features precise local editing, multilingual text rendering, and cinematic color and brightness control. Combined with Gemini 3's knowledge reasoning and search engine, it can generate highly accurate infographics.

Humanoid robot company Star Age has completed a nearly 1 billion yuan Series A+ financing round, led by Geely Capital. The company's total order value has exceeded 500 million yuan, with its largest single order in the logistics sector reaching nearly 50 million yuan. Overseas business accounts for 50% of its total revenue. This round of financing will accelerate technology iteration and commercialization.

OpenAI released the GPT-5.1-Codex-Max programming model, replacing GPT-5.1-Codex and becoming the default model in the Codex integration interface. The model outperformed Google Gemini 3 Pro in multiple programming benchmarks, such as achieving 77.9% accuracy in the SWE-Bench Verified test, leading by 1.7 percentage points. It incorporates a "compression" mechanism, allowing it to work continuously for over 24 hours, and improving token efficiency by approximately 30%.

AI video startup Luma AI has raised $900 million in Series C funding, valuing the company at $4 billion. The round was led by Humain, a subsidiary of Saudi Arabia's Public Investment Fund. The funds will be used to develop multimodal AI systems and build a 2GW intelligent computing cluster, "Project Halo".

AI music platform Suno has completed a $250 million Series C funding round, valuing the company at $2.45 billion, nearly five times its value last year. The round was led by Menlo Ventures, with participation from NVIDIA's NVentures and other institutions. Suno's core product can generate complete songs from text, and the latest v5 version significantly improves vocal naturalness and command execution accuracy.

Meta released SAM 3D and the SAM 3 model, achieving a major breakthrough in 3D reconstruction from 2D images. SAM 3D includes two models: object scene reconstruction and human pose estimation. It generates 3D labeled data from nearly one million images through an innovative data engine, outperforming existing methods in multiple benchmark tests.

Anthropic has secured a combined $15 billion investment from Microsoft and Nvidia, with Microsoft investing $5 billion and Nvidia investing $10 billion. Anthropic has committed to purchasing $30 billion worth of Azure computing power and will collaborate closely with Nvidia to optimize model performance. The Claude model will be available on Microsoft Azure, becoming a cutting-edge model accessible on all three major global cloud platforms.

AiPPT has officially launched its Agent version of the "In-Depth Research" feature and invited users to participate in internal testing. The feature simulates the workflow of a real analyst, automatically completing the entire process from in-depth web search and logical organization to generating richly illustrated research reports and PPTs through a "understanding-reflection-questioning" cycle.

Google officially released its next-generation multimodal large model, Gemini 3, which has demonstrated outstanding performance in multiple benchmark tests, ranking first in the LMARaena arena with an Elo score of 1501. The model possesses native multimodal understanding, advanced reasoning, and agent programming capabilities, supports a 1 million token context window, and significantly outperforms its predecessors and competitors in tasks such as academic reasoning, mathematics, and code development.

Google DeepMind's latest AI weather forecasting model, WeatherNext 2, is eight times faster than its predecessor and can generate accurate forecasts with a one-hour resolution. Using generative functional networks (FGN) technology, the model can complete a 15-day global weather simulation in one minute, a process that would take traditional supercomputers several hours, and can output hundreds of possible weather scenarios.

Ant Group officially launched Lingguang APP, a multi-modal general-purpose AI assistant. It generates logically clear and visual content with structured thinking, such as dynamic 3D models and interactive maps, making knowledge presentation more vivid. Users can generate personalized AI applications within 30 seconds by describing with natural language. Equipped with AGI camera technology, it analyzes the physical world through real-time video streams, supports text-to-image/video and image-to-image/video creation, and can identify objects and scenes and explain knowledge.

Elon Musk's xAI company released its latest AI conversational model, Grok 4.1, which focuses on breakthroughs in emotional intelligence and practical performance. The model topped the LMArena leaderboard with a score of 1483, 31 points higher than the second-place model, and pioneered a dual-track design of "thinking mode" and "instant mode"—even without activating deep reasoning, the basic version still surpasses the full configurations of other models.

ByteDance's Seed team released Lumine, the first general-purpose AI agent capable of completing complex tasks on an hourly scale in real-time within a 3D open world. Based on a visual language model, it unifies perception, reasoning, and action through an end-to-end architecture, relying solely on screen images and keyboard/mouse operations, without requiring in-game APIs.

JetBrains, in partnership with the Linux Foundation, launched DPAI Arena, the world's first open-source, multi-language, multi-framework AI programming agent benchmark platform. It objectively measures the efficiency improvements of AI tools in real-world software development tasks, filling a gap in the industry and supporting reproducible evaluation across diverse workflows such as patching and code review.