Search AI Products and News

February 24, 2025, the platform announced support for the new AI model Claude 3.7 Sonnet (Anthropic released Claude 3.7 Sonnet). An official X post stated that the model had been undergoing internal testing for some time, observing significant improvements in agent workflows and code generation (Perplexity Pro X post). Claude 3.7 Sonnet is described as Anthropic's most intelligent model to date.

Recently, the VITA-MLLM team announced the launch of VITA-1.5, an upgraded version based on VITA-1.0, aimed at enhancing the real-time and accuracy of multimodal interaction. VITA-1.5 not only supports English and Chinese but also achieves significant improvements in multiple performance metrics, providing users with a smoother interaction experience. In VITA-1.5, the interaction delay has been greatly reduced from 4 seconds to just 1.5 seconds, making it almost imperceptible for users during voice interactions.

Recently, IBM announced the launch of version 3.1 of its open source language model Granite, which brings many important enhancements. The new version of the model has been redesigned with a denser architecture, capable of handling up to 128,000 tokens at once. This update signifies a substantial improvement in Granite's ability to process complex text and tasks. The Granite 3.1 model has been trained on datasets in 12 languages and 116 programming languages, processing a total of 12 trillion tokens.

Recently, OpenAI has caused a stir in the video AI field as they are set to make significant upgrades to Sora, which was released in February of this year. According to reports from the Information Daily, the main goal of this upgrade is to significantly enhance Sora's performance, enabling it to generate longer and higher quality video content more quickly. Reflecting on the initial performance of Sora, there were indeed many pressing issues that needed addressing. The original version took over 10 minutes to generate a video, an efficiency level that clearly fails to meet practical application needs.

Google today announced the launch of the newly upgraded Gemini model series, which includes Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002. This update not only significantly enhances performance but also brings surprising price discounts, undoubtedly creating a buzz in the AI development community. Firstly, the most eye-catching aspect is the substantial price reduction. The usage cost of the new models has been halved, with a drop of over 50%. At the same time, performance has seen remarkable improvements.

DeepSeek V2.5 demonstrates exceptional performance in the field of artificial intelligence, particularly in code generation and chat models. Through comparative testing with GPT-4, it has achieved significant improvements across multiple metrics, including win rates, MT-Bench, and AlignBench scores. In terms of code generation capabilities, DeepSeek V2.5 achieved a HumanEval score of 89% and a LiveCodeBench score of 41%, showcasing its ability to generate high-quality, executable code.

Open-Sora has released a complete reproduction architecture scheme for Sora, which can reduce the reproduction cost by 46%. The model training input sequence length is extended to 819K patches, supporting dynamic resolution, various model structures, and multiple video compression methods. The performance improves by over 40% compared to the baseline scheme, while also reducing costs.

The performance of large language models improves as the number of instantiated agents increases. Experimental results show that multiple small models can outperform a larger model. The paper discusses the relationship between performance improvement and problem difficulty, and the author proposes two optimization strategies: progressive sampling with voting and stratified sampling with voting.

NVIDIA has launched the new RTX 500 and 1000 GPU chips, achieving a 1400% performance increase for AIGC. The RTX 500 will be the most affordable entry-level GPU with powerful global AIGC integration, attracting consumer attention due to its competitive pricing.

Stability AI Japan has recently launched two remarkable Japanese language models, named 'Japanese Stable LM3B-4E1T' and 'Japanese Stable LM Gamma7B'. These models are built upon previously released English language models and have been pre-trained with a large amount of Japanese and English data to enhance their Japanese language processing capabilities. The two models consist of 3 billion and 7 billion parameters respectively, allowing for commercial use.

Google released its new generation AI chip Cloud TPU v5e. TPU v5e provides efficient training and inference performance for medium to large scale. Jeff Dean pointed out that improving AI hardware performance has become more difficult.