According to @godofprompt, a new paper by the Ling team titled 'Every Attention Matters' introduces the Ring-linear architecture, which fundamentally changes long-context reasoning in large language models (LLMs). This architecture combines Softmax and Linear Attention, achieving a 10x reduction in inference costs while maintaining state-of-the-art accuracy on sequences up to 128,000 tokens (source: @godofprompt, Twitter, Oct 25, 2025). The paper reports a 50% increase in training efficiency and a 90% boost in inference speed, with stable reinforcement learning optimization over ultra-long sequences. These breakthroughs enable efficient scaling of LLMs for long-context tasks without the need for trillion-parameter models, opening new business opportunities in AI-driven document analysis, legal tech, and scientific research requiring extensive context windows.