Search AI Products and News

The Arc Prize foundation recently released a new test, ARC-AGI-2, designed to measure the general intelligence level of artificial intelligence (AI) models. Co-founded by renowned AI researcher François Chollet, the foundation's blog states that this new test presents a significant challenge to most leading AI models. According to the Arc Prize leaderboard, models such as OpenAI's o1-pro and DeepSeek...

Reka AI, founded by a dozen former Google DeepMind scientists, has unveiled its first open-source model: Reka Flash 3. This 21-billion parameter inference model has garnered significant attention. Despite its relatively smaller parameter count, Reka Flash 3 is a general-purpose reasoning model trained from scratch. It underwent supervised fine-tuning on synthetic and public datasets and further refinement through model-based techniques.

Tencent HunYuan recently announced the launch of five new open-source 3D generative models. Built upon Hunyuan3D-2.0, these models boast faster generation speeds, richer details, and more realistic material expression. Simultaneously, Tencent's self-developed 3D AI creation engine has been upgraded, adding features such as multi-view input, intelligent model simplification, and full format compatibility. It is now fully open to C-end users. Among these new models, the Turbo series maintains high precision and quality while optimizing the geometric generation model...

Google's new Gemini 2.0 Flash AI model is causing controversy due to its ability to easily remove watermarks from images, raising significant copyright concerns.

Frustrated by the slow processing speed of large language models on long documents? Researchers from Tsinghua University have unveiled a groundbreaking technology – the APB parallel inference framework – that dramatically accelerates processing. Benchmark tests show this technology achieves a 10x speed improvement over Flash Attention when handling ultra-long texts. With the rise of models like ChatGPT, AI's ability to process vast amounts of text (hundreds of thousands of words) has increased significantly. However, this often comes at the cost of processing speed...

Google recently launched its most cost-effective model, Gemini 2.0 Flash-Lite, now available for production use. Part of Google's Gemini family, Gemini 2.0 Flash-Lite is positioned as the most budget-friendly option. The model is offered in public preview on Google AI Studio and Vertex AI, targeting developers requiring a high value-to-cost AI solution. Its design...

On the first day of DeepSeek Open Source Week, we officially released our latest technological achievement, FlashMLA, an efficient Multi-Layer Attention decoding kernel specifically designed for NVIDIA Hopper architecture GPUs. This technology is optimized for variable-length sequence scenarios and can significantly enhance the inference performance of large models. The core technical features of FlashMLA include comprehensive support for BF16 precision and a paged key-value cache with a block size of 64.

Recently, Vectara released a report titled 'Hallucination Ranking', comparing the performance of different large language models (LLMs) in generating hallucinations while summarizing short documents. This ranking utilized Vectara's Hughes Hallucination Evaluation Model (HHEM-2.1), which is regularly updated to assess the frequency of false information introduced by these models in summaries. According to the latest data, the report highlighted key metrics such as hallucination rates, factual consistency rates, response rates, and average summary lengths of several popular models.

An innovative video generation technology, FlashVideo, has recently announced its open-source release. This framework cleverly addresses the balance between speed and quality in AI video generation through a unique two-stage design. FlashVideo employs an approach similar to LQIP (Low-Quality Image Placeholder) used in web front-end development by dividing the video generation process into two stages: preview and quality enhancement. In the preview stage, the system can quickly generate video drafts at a very low computational cost, providing users with instant feedback and significantly improving the user experience and commercial application feasibility.

Traditional video understanding models face numerous challenges when processing long videos, including the complexities of understanding the extended context. Although considerable research has been conducted to enhance video understanding capabilities, effectively overcoming the issues of low training and inference efficiency remains difficult. To address these challenges, the research team utilized HiCo technology to compress redundant parts of video information, significantly reducing computational demands while retaining key information. Specifically, HiCo achieves hierarchical compression of the video by segmenting long videos into shorter clips, thereby reducing processing time.

Beijing Zhiyuan Huazhang Technology Co., Ltd. announced on January 16, 2025, the launch of a series of new models and their availability on bigmodel.cn. Following the release of 'Zhiyuan Qingtalk' in August, the company has delved deeper into the understanding and generation of language, speech, images, and videos, introducing multimodal models such as GLM-Voice, GLM-4V, CogView, and CogVideoX.

At the 2025 CES Consumer Electronics Show, Flash announced its new overseas sub-brand Loomos and introduced the latest Loomos AI glasses. The hardware configuration and features of these glasses are largely the same as those of the domestic version, Flash AI 'Pailai Glasses', supporting 4K photo and 1080P video capture, showcasing strong imaging capabilities. Unlike the domestic version, the Loomos AI glasses are equipped with a voice assistant powered by GPT-4o, providing instant assistance to users.