《晚安咖啡》科技夜话（英文版）：Grok 4 Fast巨幅升级，微软挑战英伟达CUDA，大模型解耦推理成主流

晚安咖啡GoodNightCoffee · Nov 10, 2025 · 14m

xAI's Grok 4 Fast model significantly upgrades its context window to 2 million tokens.

Microsoft launches a toolkit converting CUDA models to ROCm compatible code, challenging NVIDIA's dominance.

Large model decoupled inference becomes a mainstream AI concept, improving efficiency and adopted by tech giants.

Google introduces its Ironwood AI chip, quadrupling performance, and strikes a multi-billion dollar deal with Anthropic.

China's MIIT accelerates the cultivation of 5G+, AI+, and Robot+ application scenarios, driving new industrialization.

The AI-driven storage chip market enters a…