Search AI Products and News

Large Language Models (LLMs) have achieved significant progress in complex reasoning tasks by combining task prompts with large-scale reinforcement learning (RL), as demonstrated by models like Deepseek-R1-Zero, which directly apply reinforcement learning to base models, showcasing strong reasoning capabilities. However, this success is difficult to replicate across different base model families, especially within the Llama series. This raises a core question: what factors lead to inconsistent performance of different base models during reinforcement learning? How does reinforcement learning perform in

The plan's advancement has triggered dissatisfaction among developers. According to VentureBeat, OpenAI's current API deprecation involves the GPT-4.5 Preview version, which is an intermediary model launched between GPT-4 and GPT-5. Multiple developers have stated that this sudden change has disrupted their development plans, forcing significant adjustments to applications already built based on this API. A developer who wished to remain anonymous commented: We understand the necessity of technological iteration, but the lack of sufficient buffer time in the schedule has brought additional pressure to development teams.

Recently, the DeepSeek team released a technical paper about their latest model, DeepSeek-V3, focusing on the challenges of scaling in large AI model training and related considerations of hardware architecture. This 14-page paper not only summarizes the experiences and lessons learned during the development of V3 by DeepSeek but also provides deep insights for future hardware design. Notably, CEO of DeepSeek, Liang Wenfeng, also participated in writing the paper.

A new Google report reveals that effective AI training for employees could unlock a £400 billion (approximately $533 billion) boost to the UK economy from AI-driven growth. Based on a UK pilot program, the report shows employees saved over 122 hours annually on administrative tasks by using AI tools. The report highlights that simplifying AI usage and providing adequate training are key to wider AI adoption. Google's European, Middle...

Google recently released a new version of its Gemma3 series, exciting many AI enthusiasts. Just a month after its initial launch, Google released a Quantization Aware Training (QAT) optimized version of Gemma3, aiming to significantly reduce memory requirements while maintaining model quality. Specifically, the QAT-optimized Gemma3 27B model reduces VRAM requirements from 54GB to 14.1GB, meaning users can now run it on a single NVIDIA RTX 3090.

Hugging Face recently released its top model leaderboard for the second week of April 2025, encompassing various modalities such as text generation, image generation, and video generation. This highlights the rapid iteration and diverse applications of AI technology. According to AIbase, the models featured in this week's leaderboard not only showcase the innovative vitality of the open-source community but also reflect technical trends such as low-precision training and multi-modal generation. Below is an analysis of the leaderboard highlights, with professional insights provided by the AIbase editorial team. Text Generation Models: Balancing Efficiency and Specialization

Kingsoft Cloud announced that its StarStream training and inference platform has fully integrated with Zhipu's GLM series inference models, becoming one of the first platforms to integrate this series. This move marks Kingsoft Cloud's further expansion in the AI field, providing users with more efficient, intelligent, and cost-effective model services.