OpenRouter publishes a global large language model analysis report based on 100 trillion token data: open-source model usage increased to 30% within a year, the share of Chinese open-source models leaped from 1.2% to 30%, and medium-scale (15B-70B) models are rising; role-playing and programming are the primary application scenarios, with reasoning models rapidly rising to become mainstream (over 50% share) within a year; the 'Glass Slipper Effect' shows extremely high user stickiness after finding a core workload match, emphasizing that commercial success relies more on user retention than short-term spikes; trends of multi-model ecosystems, multi-scenario adaptation, new cost/performance balance, and globalization are significant, with Agentized reasoning becoming the new competitive frontier.

Alibaba's Qwen team releases SAPO (Soft Adaptive Policy Optimization), a reinforcement learning optimization method for large models. It introduces a temperature-controlled soft-gating mechanism, improving convergence and sample efficiency during large language model reinforcement training, balancing stability and breakthrough learning capabilities. This lays an innovative foundation for RLHF training of domestic large models.

Google DeepMind CEO Demis Hassabis stated that continuing to scale up AI models and data is key to achieving Artificial General Intelligence (AGI), but relying solely on 'scaling laws' is insufficient for mastering all intelligent capabilities. Truly advancing toward AGI may require 1-2 fundamental paradigm breakthroughs. Industry figures like Yann LeCun have expressed skepticism about the logic of 'piling on data and compute.'

Latest rankings from international benchmarks like LMArena show several domestic large models, including Baidu's Wenxin, ranking among the global top in areas like text understanding, creative writing, and complex long-form reasoning. Domestic vendors have entered the top tier of AI competition, achieving breakthroughs in practical, multi-domain AI capabilities, leading to increased international influence for Chinese large models.

Amid the global AI training and inference boom, PC DRAM and SSD supplies fell short of demand in the second half of 2025. Prices surged as major manufacturers like Micron prioritized supply to the AI server market, squeezing the traditional consumer market. Industry analysis suggests that AI infrastructure investment has raised overall storage hardware costs, affecting the pace of end-device upgrades.

A Bernstein report predicts Nvidia's market share in the Chinese AI chip market will fall to 8% by 2026, with Huawei and other local Chinese AI chip manufacturers significantly increasing their shares. This marks breakthroughs in the domestic computing ecosystem across fabrication, algorithms, and equipment, further weakening the effectiveness of US export restrictions.

Multiple media reports indicate OpenAI recently decided to pause all new product development and activities outside of ChatGPT, concentrating all R&D resources on fixing core ChatGPT experience and long-term stability issues. This signals that large model platform operations have entered a new phase post-GPT-4o, characterized by high productization and reliability prioritization.

Recent research reveals that 62% of mainstream large models (including ChatGPT) cannot defend against standard input bypasses using creative text like poetry and rhyme to reject non-compliant requests. This exposes severe blind spots in text understanding security for large models. Future jailbreak offense and defense will heavily rely on model security iteration.

Driven by AI infrastructure demand, Databricks expects its 2024 revenue to rise to $4.1 billion. The company's target valuation for a new external funding round is as high as $134 billion, reflecting that the penetration and growth rate of AI enterprise data platforms and large model tools are surpassing traditional SaaS, with industry investment and financing scales repeatedly hitting new highs.

According to OpenRouter and major platform statistics, the usage rate of Chinese open-source large models like Qwen, DeepSeek, Kimi, and MiniMax surged to 30% globally in 2025. Individual models like DeepSeek lead global token usage ahead of Meta Llama and Qwen. This indicates that the open-source ecosystem has become the dominant force in the international large model market fragmentation, with Chinese vendors continuously advancing in areas like multilingual support, programming, and reasoning assistance.