China's AI "Qwen3.8-Max" Surpasses GPT-5.6, Open-Source Release Coming Next Week

What you'll learn in this article

• The capabilities and features of Alibaba's "Qwen3.8-Max"
• Why it outperforms GPT-5.6 and Fable 5
• How it runs at low cost despite having 2.4 trillion parameters
• The impact on Japanese businesses

Alibaba Officially Releases "Qwen3.8-Max"

On August 3, 2026, China's Alibaba Cloud unveiled its latest AI model, "Qwen3.8-Max."

Despite its massive scale of 2.4 trillion parameters, the model is designed to operate efficiently by activating only 9.5 billion parameters at a time.

It surpasses cutting-edge models such as Claude Fable 5 and GPT-5.6 Sol on certain benchmarks, and an open-source release is planned for sometime next week.

What Makes It Stand Out

On "Terminal Bench 2.1," a benchmark measuring coding ability, it outperformed Claude Fable 5, and on "SWE-bench Pro," it surpassed GPT-5.6 Sol.

On "PaperBench," which measures research paper comprehension, it scored 93.0, exceeding GPT-5.6 Sol (90.5) and Fable 5 (88.8).

How "MoE" Lets 2.4 Trillion Parameters Run on Just 9.5 Billion

Qwen3.8-Max employs a structure called "MoE (Mixture of Experts)."

Think of it like having a team of 2.4 trillion specialists — when a question comes in, only the 9.5 billion specialists needed to answer it are activated.

This mechanism dramatically reduces computational costs and response times, enabling a low price point of $2 per million input tokens and $6 per million output tokens.

The same approach is used in DeepSeek's V4 series and Mistral's Mixtral, making it a key technology that gives Chinese AI players a competitive edge in cost competition.

What Open-Sourcing Means

Alibaba announced that the model weights for Qwen3.8-Max will be published on Hugging Face and ModelScope sometime next week.

Once open-sourced, developers around the world will be able to deploy it on their own servers and use it for free.

Developers who have already migrated to Qwen 3.7 Max have reported an "87% reduction in API costs."

The Rise of Chinese AI

As of July 2026, approximately 82% of the top 10 endpoints on OpenRouter — a popular AI access service — by monthly ranking are occupied by Chinese-developed models.

The key players are as follows:

• DeepSeek: An exceptionally low price of $0.14 per million input tokens
• Alibaba Qwen: The most widely adopted open model on Hugging Face
• Kimi: Announced "Kimi K3" with 2.8 trillion parameters
• GLM: Strong capabilities in coding assistance

These models are closing in on Western state-of-the-art performance while being dramatically cheaper — a dual advantage.

Impact on Japanese Businesses

The open-sourcing of Qwen3.8-Max presents both opportunities and risks for Japanese businesses.

The opportunity lies in cost reduction for workloads that process large volumes of tokens, such as customer support chatbots and internal document summarization.

On the other hand, when using AI models provided by Chinese companies, careful attention must be paid to data sovereignty and cross-border data transfer regulations.

In highly sensitive industries such as finance and healthcare, the location of data storage and access permissions must be carefully considered.

Summary

• Alibaba has released "Qwen3.8-Max" with 2.4 trillion parameters, with an open-source release planned for next week
• Outperforms GPT-5.6 Sol and Claude Fable 5 on certain benchmarks
• Achieves low cost and fast responses by activating only 9.5 billion parameters via MoE architecture
• Chinese AI players hold an 82% share of the global market
• A cost-reduction opportunity for Japanese businesses, but data sovereignty considerations are essential

---

This article is a cross-post from AI Friends.