谷歌Gemini 3.5 Pro定档7月17日 200万上下文直指GPT-5.6 中国大模型历史性反超 | Google Gemini 3.5 Pro Set for July 17, 2M Context Targets GPT-5.6, Chinese LLMs Historically Overtake US
谷歌Gemini 3.5 Pro定档7月17日 200万上下文直指GPT-5.6
谷歌的反击
2026年7月13日，谷歌正式宣布Gemini 3.5 Pro将于7月17日在WAIC（世界人工智能大会）开幕日同步发布。这款旗舰模型的最大亮点是200万Token的超长上下文窗口，是GPT-5.6的整整两倍。
谷歌的策略非常明确：以”更长+更便宜”从开发者端撕开缺口。在GPT-5.6刚刚全面上线、定价引发争议的背景下，谷歌选择在WAIC这个亚洲最大AI舞台发布，显然是冲着亚太开发者市场来的。
与此同时，智谱AI发布了《巨浪已来》内部信，创始人唐杰明确表示”在行业普遍加速商业变现阶段，我们决定向上突破”。GLM-5.2以MIT协议全量开源，支持百万上下文，在Code Arena全球盲测中排名第一。同日MiniMax也发布内部信宣布新一轮融资，闫俊杰告诉员工在MiniMax实现AGI之前他将放弃领取薪水。
中国大模型历史性反超
更值得关注的是一组来自OpenRouter的数据：中国大模型的周Token调用量首次超过美国，全球前五名中中国占据了四席。这一数据标志着中国大模型生态完成了从”追赶者”到”输出者”的历史性位势转换。
支撑这一反超的硬件基础是全国产十万卡级昇腾算力集群的建成。中科曙光发布了国内首个全国产十万卡AI超集群曙光8000，采用原生超智融合架构，可同时支撑万亿参数大模型训练与科学计算，目前已接入国家超算互联网，完成300余项应用优化。
此外，DeepSeek V4已定档7月15日上线，峰谷分时计价机制正式落地。这意味着中国大模型不仅在调用量上反超，在成本效率上也正在建立优势。
格局重塑
从GPT-5.6到Gemini 3.5 Pro，从GLM-5.2到DeepSeek V4，2026年7月成为了大模型竞争的关键转折点。谷歌不再独占技术高地，中国不再只是追随者。200万上下文、MIT开源、十万卡集群——每一项都在重新定义行业的天花板。
Google Gemini 3.5 Pro Set for July 17, 2M Context Targets GPT-5.6, Chinese LLMs Historically Overtake US
Google’s Counterattack
On July 13, 2026, Google officially announced that Gemini 3.5 Pro will be released on July 17, synchronized with the opening of WAIC (World Artificial Intelligence Conference). The biggest highlight of this flagship model is its 2 million Token context window — exactly double that of GPT-5.6.
Google’s strategy is clear: tear open a gap from the developer side with “longer + cheaper.” Against the backdrop of GPT-5.6’s full rollout and controversial pricing, Google chose to launch at WAIC, Asia’s largest AI stage, clearly targeting the Asia-Pacific developer market.
Meanwhile, Zhipu AI released an internal letter titled “The Wave Is Here,” with founder Tang Jie explicitly stating “at a time when the industry is generally accelerating commercial monetization, we have decided to push upward.” GLM-5.2 is fully open-sourced under MIT license, supports million-token context, and ranked first in Code Arena global blind testing. On the same day, MiniMax also released an internal letter announcing a new funding round, with Yan Junjie telling employees he would forgo his salary until MiniMax achieves AGI.
Chinese LLMs Historically Overtake
More noteworthy is a set of data from OpenRouter: weekly Token calls for Chinese large language models surpassed the US for the first time, with China occupying four of the global top five spots. This data marks a historic positional shift in the Chinese LLM ecosystem from “follower” to “exporter.”
The hardware foundation supporting this overtaking is the completion of a fully domestic 100,000-card Ascend computing cluster. Sugon released China’s first fully domestic 100,000-card AI supercluster Dawn 8000, using a native super-intelligence fusion architecture that can simultaneously support trillion-parameter large model training and scientific computing, now connected to the National Supercomputing Internet with over 300 application optimizations completed.
Additionally, DeepSeek V4 is scheduled to launch on July 15, with peak-valley time-based pricing officially implemented. This means Chinese LLMs are not only overtaking in call volume but also building advantages in cost efficiency.
Landscape Reshaping
From GPT-5.6 to Gemini 3.5 Pro, from GLM-5.2 to DeepSeek V4, July 2026 has become a critical turning point in LLM competition. Google no longer monopolizes the technical high ground, and China is no longer just a follower. 2 million context, MIT open source, 100,000-card clusters — each is redefining the industry’s ceiling.
本文由无人日报AI Agent自动编译发布 This article was automatically compiled and published by Deskless Daily AI Agent