As the Mid-Autumn Festival and National Day approach, the AI industry is also about to witness a major wave — at least 10 large models are expected to be launched around the holidays, both domestically and internationally.

Abroad: OpenAI completes the GPT-6 series, Anthropic skips 5.1

Although American companies don't usually release products before holidays, this timing is quite coincidental: OpenAI has a developer conference at the end of the month, and the "big product" previously announced for this week has been postponed to next week, falling near the Chinese holiday. OpenAI's new products include at least GPT-6 Sol, GPT-6 Terra, and GPT-6 Luna, forming a full series upgrade with the already released GPT-6 Astra — after all, Astra is relatively expensive and not suitable for daily heavy use.

Although Anthropic has called for slowing down AI research, it hasn't stopped itself. Its next batch of products even skipped 5.1 and directly released Opus 5.2 and Fable 5.2. Mythos 5.2, which will not be widely publicly available, may also be released.

Google may launch Gemini 4.0 Pro and 4.0 Flash, which is seen as a crucial battle for whether it can remain in the "Big Three." Previous leaked results were very impressive, but their credibility is generally low. Grok 4.7 from Elon Musk's xAI was originally scheduled to be released around September 12th, but it has been postponed. Various signs suggest its performance may not bring much surprise. Microsoft, Meta, and Amazon have not shown any signs of releasing major products recently.

Domestic: Step 5 is now live, K3.1, Qwen4, V4.1 Pro are waiting in line

Domestically, the first to act today is Step 5 Preview by Jieque Star, which has officially launched with a major promotional activity. The next more certain one is Kimi K3.1 — the official recently released a string of mysterious numbers, which netizens interpreted as hints for K3.1, but then deleted the post; considering that K3 has been released for several months, K3.1 needs to be launched as soon as possible to maintain its ranking position.

The Qwen4 series from Alibaba's Tongyi is also worth looking forward to. It has already demonstrated a new architecture on Qwen 3.8 Flash, and the new small model that can be deployed locally, such as Qwen4 27B, has received high praise. Xiaomi has publicly released the new Mimo model and client preview, which should be the Mimo V2.6 series. Recently, it has also been conducting RL post-training live broadcasts, with costs reaching 2.4 million USD. There are rumors that GLM-5.5 from Zhipu will be upgraded to more than 1 trillion parameters; MiniMax M3.1 is also rumored to be upgraded to 1 trillion parameters and lay the foundation for the subsequent 3.1 Pro to aim for 3 trillion parameters. After HY3 and HY4, Tencent also needs a large parameter HY4 Pro as its flagship.

The most anticipated one remains DeepSeek: the V4.1 Flash released on the 10th received widespread praise, with significant architectural upgrades but no major name changes. The Pro series should be V4.1 Pro — if it also adopts the CED architecture, the parameter count may double from 1.6 trillion in V4 Pro to 3 trillion. Whether it can redeem the previous humiliation of V4 Pro-0813 is the biggest highlight.