Skip to main content

AI Model Blitz: 10+ New Releases Set to Land Around the Holidays

A Holiday Flood of AI Models

As the Mid-Autumn Festival and National Day approach, the AI industry is preparing for an unprecedented wave of releases. At least 10 major models—both from overseas and domestic players—are expected to debut around the holidays. For anyone tracking AI, it's like waiting for a fireworks show.

Overseas: OpenAI's Full House, Anthropic's Skip

OpenAI has a developer conference at the end of the month, and a previously teased "big product" has been pushed to next week, conveniently near the Chinese holiday. The company is expected to unveil GPT-6 Sol, Terra, and Luna, completing the GPT-6 family alongside the already-released Astra. Astra, while powerful, is pricey for everyday use, so these new siblings should offer more practical options.

Anthropic, despite preaching caution on AI research, isn't slowing down. It's skipping version 5.1 entirely and going straight to Opus 5.2 and Fable 5.2. A limited-release Mythos 5.2 may also appear.

Google might counter with Gemini 4.0 Pro and 4.0 Flash—a crucial test for its standing in the "Big Three." Early leaks look impressive, though skepticism remains. Meanwhile, xAI's Grok 4.7, originally slated for mid-September, has been delayed, and insiders suggest it may not wow. Microsoft, Meta, and Amazon are quiet for now.

Domestic: Step 5 Leads, DeepSeek V4.1 Pro Awaits

On the home front, Step 5 Preview from Jieque Star launched today with a major promotional push. Next up is likely Kimi K3.1—the company recently posted a cryptic number string that fans decoded as a K3.1 hint, only to delete it. With K3 out for months, an update is overdue to keep its ranking.

Alibaba's Qwen4 series is also highly anticipated. The new architecture already showed promise on Qwen 3.8 Flash, and a locally deployable small model like Qwen4 27B has earned praise. Xiaomi has teased its new Mimo V2.6 series and client preview, even live-streaming RL post-training that cost $2.4 million. Rumors swirl that Zhipu's GLM-5.5 will surpass 1 trillion parameters, while MiniMax M3.1 may also hit 1 trillion, setting the stage for a 3-trillion-parameter 3.1 Pro. Tencent, after HY3 and HY4, needs a flagship HY4 Pro.

But the star of the show remains DeepSeek. The V4.1 Flash, released on the 10th, won widespread acclaim for its architectural upgrades without a name change. The Pro version, V4.1 Pro, is expected to adopt the CED architecture, potentially doubling parameters from 1.6 trillion to 3 trillion. Will it redeem the misstep of V4 Pro-0813? That's the biggest question.

Key Points

  • Overseas: OpenAI's GPT-6 Sol, Terra, Luna; Anthropic's Opus 5.2 and Fable 5.2; Google's Gemini 4.0 Pro/Flash; xAI's delayed Grok 4.7.
  • Domestic: Step 5 Preview live; Kimi K3.1 imminent; Qwen4, Xiaomi Mimo V2.6, Zhipu GLM-5.5, MiniMax M3.1, Tencent HY4 Pro in pipeline.
  • DeepSeek V4.1 Pro could double parameters to 3 trillion, aiming to make amends for V4 Pro-0813.
  • The holiday season will be a stress test for AI companies' ability to deliver on hype.