Workflow
Eve(Grok 4语音助手)
icon
Search documents
Grok 4长流程工作应用潜力初显 带动AI Infra与算力需求
智通财经网· 2025-07-12 07:50
Core Viewpoint - The release of Grok 4 by XAI demonstrates significant advancements in reasoning capabilities for professional disciplines and complex tasks, indicating potential applications in high-value scenarios and driving demand for AI infrastructure and computing power [1][2]. Group 1: Product Release and Pricing - Grok 4 has been officially launched, featuring two versions: Grok 4 and Grok 4 Heavy, with enhanced performance in professional tasks [2]. - The pricing for the B-end API is set at $3 per million tokens for input and $15 per million tokens for output, approximately 50% higher than the previous version [2]. - C-end users can access Grok 4 for a subscription fee of $30 per month, while the high-performance Grok 4 Heavy version costs $300 per month [2]. Group 2: Performance Enhancements - Grok 4 significantly outperforms previous state-of-the-art models in reasoning tasks, achieving a 26.9% accuracy rate without tools and 41.0% with tools on the Humanity's Last Exam (HLE) test set, with potential to reach 50.7% through increased reinforcement learning (RL) computation [3]. - In the Vending-Bench test, Grok 4 scored twice as high as the second-place model, Claude Opus 4, indicating its capability in solving complex real-world problems [3]. - Grok 4 Heavy excelled in several academic knowledge tests, achieving near-perfect scores in AIME25 and HMMT25 [3]. Group 3: Computational Demand and Technical Innovations - The training volume for Grok 4 has increased by 100 times compared to Grok 2, and the computational load for post-training reinforcement learning has increased tenfold compared to Grok 3 [4]. - Grok 4 Heavy has validated the effectiveness of increased RL computation in enhancing model performance, demonstrating superior cost-effectiveness in reasoning compared to all previous models [4]. - Key engineering innovations include the importance of tool usage in improving reasoning performance and the development of reliable reward signal schemes in post-training reinforcement learning [4]. Group 4: Future Developments and Multimodal Capabilities - The new voice assistant, Eve, has reduced conversation latency by 50% and increased daily user engagement by tenfold, showcasing advanced conversational abilities [5]. - There are plans to enhance visual understanding and generation capabilities in upcoming updates, with a focus on multimodal intelligence [5]. - Future releases include a code model in August, a multimodal agent in September, and a video generation model in October [5].