Search AI Products and News

OpenAI announced that GPT-5 will be launched soon, but clarified that it is not the experimental model that won the International Mathematical Olympiad. The CEO stated that the winning model used new technology and has mathematical capabilities far beyond current levels, while GPT-5, though expected to provide a surprising experience, will have different mathematical abilities. The community is热议 a suspected GPT-5 test model appearing on GitHub. OpenAI emphasized the need to distinguish between the actual capabilities of different models, providing clear guidance for market expectations. The AI field continues to closely watch the release of GPT-5 and technological breakthroughs. (140 characters)

FotographerAI introduces the revolutionary image generation framework ZenCtrl, which enables high-fidelity image generation with multiple perspectives and scenes from a single input image using GenAI technology. This framework breaks through three major pain points of traditional image generation: no need for fine-tuning, maintaining subject consistency, and supporting fine-grained control. Its modular toolkit integrates the entire workflow from preprocessing to postprocessing, built upon and optimized from OmniControl, with particular emphasis on enhancing subject retention capabilities. It supports diverse scenarios such as advertising creativity and virtual try-on, and future developments are also planned.

In 2025, open-source large models will show three major trends: 1) MoE architecture becomes mainstream, with DeepSeek-V3 (67.1 billion parameters) and Qwen3-235B (235 billion parameters) each having unique designs in their expert systems; 2) small models break through performance bottlenecks, with SmolLM3-3B adopting position-encoding-free technology, and Qwen3-4B achieving lightweight efficiency; 3) models show significant differentiation, with Llama3.2 focusing on general tasks, while Kimi-K2 (1 trillion parameters) excels in complex reasoning.

Recent research questions the mathematical ability of Alibaba's Qwen2.5 model, suggesting that its high scores may stem from data memorization rather than genuine reasoning. Tests show that Qwen2.5's performance drops sharply on previously unseen clean benchmarks; however, it can accurately reconstruct 60% of known questions, indicating training data contamination. The completion rate in the new LiveMathBench test dropped to zero. The study warns that contaminated benchmarks can mislead AI progress evaluation and emphasizes the need for clean benchmarks and multi-model validation. This finding raises doubts about the effectiveness of current AI evaluation methods and highlights the issue.

Stability-AI's SPAR3D model revolutionizes 3D reconstruction from single images in 0.7s, combining regression and generative modeling via a two-stage point sampling and meshing architecture. It outperforms traditional methods in geometric accuracy and texture quality on GSO and OmniObject3D datasets. Open-sourced.....

French AI startup Mistral is negotiating $1B funding, with valuation soaring to €5.8B within a year. Founded by ex-Meta/DeepMind staff, its open-source Mistral7B outperforms Llama2. Rapid funding (€500M in 2023) fuels global AI expansion.....

Lightricks releases open-source LTX-Video13B, a 13B-parameter video generation model with multi-scale rendering, achieving 30x faster speeds. It runs on consumer GPUs, supports 1216×704 real-time generation, and offers text/image/video-to-video modes. The model enhances coherence and detail, enabling keyframe control and style transfer. Free for SMEs, it includes training tools and optimized versions to democratize AI video creation.....

Yang Jianchao, the head of ByteDance's Visual Large Model team, announced a temporary leave due to family reasons, with Zhou Chang, former technical leader of Alibaba's Tongyi Qianwen, taking over. This personnel change occurred during a period of adjustment in ByteDance's AI department, sparking concerns about the stability of the technical roadmap. Yang Jianchao's work information remains in the internal system, and Zhou Chang will lead the global Seed team to continue research on visual multimodal generation. The company emphasized its continued investment in basic research and hopes that the new leader will bring innovative energy. This change highlights the importance of balancing work and health in the fast-paced tech industry.

NVIDIA's Canary-Qwen-2.5B sets a 5.63% WER record on Hugging Face OpenASR. This CC-BY licensed model combines FastConformer encoder with Qwen3-1.7B LLM decoder for efficient speech-to-text and NLP. Supports multi-GPU deployment for cloud/edge applications.....

Lanxin Tech unveiled 'Lanxin Jiugong AI Energy Model' at the Chain Expo, featuring core technologies like a time-series prediction engine with 90% accuracy and an AI agent engine for real-time monitoring and strategy generation, now applied in Guangdong, Shandong, and Zhejiang.....

Mistral AI's chatbot Le Chat has received a major update with five core features now available: 1) Deep Research mode that intelligently breaks down complex questions and generates structured reports; 2) Voice input feature based on the Voxtral model for natural conversation; 3) Thinking mode using the Magistral model to handle complex reasoning; 4) A text-to-image modification feature developed in collaboration with the Black Forest Lab; 5) New project management tools to organize conversations and files. These features are now available on both web and mobile platforms, significantly enhancing AI capabilities.

MirageLSD, the world's first AI real-time video conversion model with 40ms latency, enables instant scene/outfit changes via gestures. Applied in gaming/livestreaming, its LSD model uses Diffusion Forcing to eliminate long-generation errors.....

Google launched Veo3 API for text-to-video & audio generation, but costs are high ($0.75/s for 720p). Used in gaming cinematics & 3D modeling. Faster 'Veo3Fast' mode not yet API-accessible. Requires Gemini API & Google Cloud subscription.....

AI coding tool Windsurf announced the re-launch of Anthropic's Claude Sonnet 4 model, offering Pro users a monthly quota of 250 API calls (2x credit consumption). The model is known for its 72.7% performance on the SWE-bench test, supports a 200K token context window, and enables code generation and complex refactoring features. Previously, due to Anthropic's restrictions on direct access, Windsurf introduced a BYOK solution. This partnership restoration is being

Baidu AI Assistant introduces 'Video Call' feature, enabling real-time video chats via Baidu APP. It covers four scenarios: object/insect identification, styling advice, pet behavior analysis, and emotional support, with dialect recognition for elderly users. This upgrade transforms AI from a tool to a smart life companion.....

Huawei, Yunnan Communications Investment, and Chang'an University launched the 'Green and Beautiful Corridor·Transportation Model' to drive digital transformation in transport. It features an 84% accurate cognitive model, AI computing centers, and 35 edge nodes, enhancing efficiency and safety in construction and management.....

Recently, the ByteDance Seed team collaborated with the University of Hong Kong and Fudan University to introduce an innovative reinforcement learning training method called POLARIS. This method successfully enhances the mathematical reasoning capabilities of small models to levels comparable to those of large models through a carefully designed Scaling RL strategy, offering a new approach for optimizing small models in the field of artificial intelligence. Experimental results show that the 4 billion parameter open-source model Qwen3-4B trained using POLARIS achieved remarkable performance on AIME25 and AIME24 mathematical tests.

TRAE.ai launches Kimi-K2 model and Grok-4(Beta). Kimi-K2 excels in code/math with MoE architecture, rivaling GPT-4.1. Easy 3-step access. International version adds Grok-4(Beta) testing alongside Claude, Gemini, GPT.....

Unsloth AI quantized Moonshot AI's 1T-parameter Kimi K2 model to 1.8bit, reducing size by 80% to 245GB while maintaining performance. The MoE-based model excels in coding and reasoning, now deployable on 512GB M3Ultra devices, lowering costs. This advancement positions Kimi K2 as a GPT-4.1 competitor, benefiting SMEs and boosting open-source AI adoption in education/healthcare.....