xAI 推出 Grok 4.20 Beta，四智能體團隊

🔑 增強重點摘要

• •Grok 4.20 Beta was released on February 17, 2026, as a public beta accessible to X Premium+ and SuperGrok subscribers via manual selection in the model menu[1][2].
• •The model achieves 95% accuracy on MMLU-Pro benchmark, up from 85% in Grok 3, with up to 10× faster response times and advanced image/video understanding[1].
• •Trained on the Colossus supercluster using 200,000 GPUs, it features approximately 3 trillion parameters in a Mixture-of-Experts (MoE) architecture and supports a 256K+ context window expandable to 2M[2][5].
• •Grok 4.20 Heavy mode scales to 16 agents for tougher tasks, and Elon Musk announced it on February 18, 2026, highlighting its integration with Tesla's ecosystem[1][3].
• •It excels in real-world benchmarks like Alpha Arena (profitable trading) and supports use cases such as medical document analysis and engineering diagram review via multimodal inputs[2][4].

🛠️ 技術深入

• •Native multi-agent system uses four specialized replicas of a ~3T-parameter MoE model that collaborate in real-time on complex queries, with adaptive activation (Fast/Expert/Heavy modes up to 16 agents)[1][5].
• •Trained on Colossus supercluster with 200k GPUs; supports 256K+ context window (up to 2M), native multimodal processing for text/image/video[2][5].
• •Internal agent traces are optional/compressed; reasoning tokens billed per API norms, with cross-agent error-checking to reduce hallucinations and 'rapid learning' from user feedback for weekly upgrades[1][5].
• •Enhanced step-back reasoning, real-time X data integration, and superior STEM/coding performance[1][4].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線