Summary of Key Points
Alibaba has released a new generation of foundational large model, Qwen3.8-Max, with a parameter scale of 2.4 trillion, marking the first time its flagship Max series has been open-sourced. This model achieves "large parameters + low latency" through a Sparse Mixed Expert (MoE) architecture, supporting contexts of up to 1 million tokens and multi-modal visual capabilities, placing it among the top performers globally. Alibaba's shift from closed-source to open-source is a strategic move in response to the challenge posed by domestic open-source models like DeepSeek. By making the technology available for free, Alibaba aims to lower the barriers for developers and gain a voice in the AI Agent ecosystem. Additionally, leveraging its comprehensive technology stack (cloud, chips, models), it seeks to create a competitive advantage in terms of computing power costs and ecosystem layout, thereby collaborating with domestic players like DeepSeek to drive the transition of global small and medium-sized enterprises' AI infrastructure towards an open-source framework.
Detailed Analysis
1. The New Model: Large, Intelligent, and Fast
The core technical strengths of Qwen3.8-Max can be summarized in three key terms: large capacity, fast response times, and advanced capabilities:
- 2.4 trillion parameters without being cumbersome: It uses the MoE architecture, which is like having 100 experts available but only activating the most relevant 10 (about 95 billion parameters) at a time. This approach retains the intelligence of large models while avoiding the slow performance and high costs associated with fully activated parameters.
- 1 million token context = handling extensive text: With 1 million tokens, the model can process content equivalent to around 700,000 to 800,000 words (e.g., a novel like "The Three-Body Problem"), allowing it to remember long pieces of information. For example, if you upload a technical manual, the model can provide detailed answers based on the entire document without needing to feed data segment by segment.
- Multi-modal vision: Previous large models could only handle text, but Qwen3.8-Max can also analyze images (e.g., recognize chart data) and videos (e.g., extract key visual information), making it suitable for more complex tasks such as design and scientific research analysis.
Alibaba claims that its performance in areas like programming, office work, and scientific research is now on par with top models like GPT-4 and Claude 3.
2. A Major Change in Business Strategy: From Closed-Source to Open-Source
The most surprising aspect of this announcement is the open-sourcing of the flagship Max model. Previously, the Max series was closed-source, and users could only access it through APIs at a per-use fee. Now, the model weights are being made available for free, allowing developers to deploy it themselves. The reasons behind this change are:
- Pressure from Open-Source Models: DeepSeek's V4 Flash model, with its superior performance and lower costs (much lower inference costs), has seen widespread adoption, demonstrating that open-source models with low prices can compete with closed-source ones. High API fees for closed-source models (e.g., several dollars per use) are becoming unsustainable as users realize the cost-saving benefits of open-source alternatives.
- Strategic Move to Gain Control in the AI Agent Ecosystem: AI Agents represent the future trend, and these applications require flexible deployment that is not dependent on a single company's APIs. By making its models open-source, Alibaba aims to dominate this ecosystem, similar to how Android has become the standard platform for mobile devices.
- Alibaba's Financial Independence: Jack Ma stated that pure model companies struggle to make money from open-sourcing, but Alibaba, with its comprehensive technology stack (cloud, chips, applications), can still generate revenue from other sources even if the models are free. Developers will need Alibaba Cloud computing resources and Pingtouge chips to deploy the models.
3. The Importance of Affordability After Open-Sourcing
Open-sourcing does not mean free access; running the model requires computing power, especially for AI Agents that need to be called frequently (e.g., during coding or data retrieval). Qwen3.8-Max's advantage lies in its extremely low cost:
- Cost Comparison: The cost per million tokens is only $6, which is 1/8 of the industry benchmark. For example, the Fable 5 model previously cost $50, while Alibaba’s model costs only 12% of that. While a few dollars may not seem significant for a single interaction, for Agents that need to be executed dozens or hundreds of times, the cost difference can significantly impact profitability.
- How Achieved: This is due to a combination of software and hardware optimizations. In addition to the MoE architecture, Alibaba uses self-developed AI chips, CPUs, and network cards from Pingtouge, as well as the Lingjun super nodes from Alibaba Cloud, which are capable of efficiently running models with over 2 trillion parameters (the first of their kind in China). This synergy reduces computing costs significantly.
- Contrast with DeepSeek: News reports indicate that DeepSeek V4 Flash experienced server failures due to high traffic loads because it lacks its own infrastructure. Alibaba, with its cloud services, is better prepared to handle such demands.
4. The Complementary Role of Domestic Models
This release is not a solo effort by Alibaba; it represents a collaboration with domestic open-source models like DeepSeek:
- Alibaba as the Foundation: Qwen3.8-Max provides a multi-modal, long-contextual foundation suitable for building complex AI Agents.
- DeepSeek as the Efficient Engine: Its low inference costs make it ideal for executing tasks such as rapid code execution and data retrieval.
- Combined Strength: This combination offers developers a flexible and cost-effective solution that can compete with foreign closed-source models like GPT-4, making domestic open-source options more attractive.
5. Global Impact
The progress of Chinese open-source models (in terms of inference efficiency and cost control) is driving the transition of global small and medium-sized enterprises' AI infrastructure towards open-source solutions:
- Why Choose Open-Source?: Closed-source models have expensive APIs and rely on a single vendor, which can be problematic in case of network outages. Open-source models can be deployed locally or across multiple clouds, offering lower costs and greater flexibility.
- Leading Companies Also Turning to Open-Source: Even giants like Alibaba are open-sourcing their flagship models, indicating that open-source is the future trend. It’s more profitable to control the ecosystem rather than relying on closed-source models for profits.
This shift signifies a global transition from a closed-source-dominated AI landscape to one where open-source models play a dominant role, with Chinese models playing a key role in this transformation.
In Conclusion
Alibaba's release of Qwen3.8-Max and its open-sourcing strategy demonstrate its commitment to leveraging technical advantages and a comprehensive technology stack to gain a foothold in the AI Agent era. By collaborating with domestic open-source models like DeepSeek, Alibaba aims to compete with foreign giants while also building a loyal developer base through open-source initiatives. For users, this means access to cheaper and more powerful AI tools in the future. For the industry, Chinese open-source models are becoming a new choice for global AI infrastructure development.