Alibaba and DeepSeek Drive China’s AI Models Toward Lower Costs

China’s artificial intelligence industry is entering a new phase where efficiency, affordability, and accessibility are becoming just as important as model size. Rather than competing only to build the largest large language models (LLMs), leading AI companies are now focusing on reducing inference costs while maintaining high performance.

Alibaba, DeepSeek, and Moonshot AI are leading this shift with powerful AI models that combine advanced architectures with competitive pricing. Alibaba recently introduced Qwen3.8-Max, its most advanced AI model so far, while DeepSeek has attracted industry attention with its V4-Flash model, offering one of the lowest inference costs currently available. Meanwhile, Moonshot AI continues to compete with its large-scale Kimi K3 model.

The competition highlights a broader trend in AI development: delivering strong performance while lowering operational costs for developers and enterprises.

Alibaba Launches Its Largest AI Model Yet

Alibaba has expanded its Qwen model family with the launch of Qwen3.8-Max, the company’s largest AI model to date. Built using a Mixture-of-Experts (MoE) architecture, the model contains approximately 2.4 trillion parameters.

Unlike traditional dense AI models that activate every parameter during inference, the Mixture-of-Experts approach activates only a small portion of the model for each request. Alibaba reports that only around 95 billion parameters are active during inference, significantly reducing computational requirements, improving response speed, and lowering operational costs.

Qwen3.8-Max is designed as a multimodal AI model capable of processing text, images, and video. It also supports an impressive one-million-token context window, allowing users to work with extremely large documents, lengthy conversations, and complex reasoning tasks without losing context.

Alibaba further demonstrated the model’s capabilities by stating that Qwen3.8-Max successfully completed a software engineering project over a continuous 16-day period, showcasing its ability to manage long-running development tasks.

Competing Directly With Moonshot AI’s Kimi K3

Alibaba’s latest release positions Qwen3.8-Max in direct competition with Moonshot AI’s Kimi K3, another large Mixture-of-Experts model introduced earlier.

According to available specifications, Kimi K3 contains approximately 2.8 trillion total parameters, with around 104 billion active parameters during inference. Although Kimi K3 is slightly larger overall, both models rely on sparse activation to improve efficiency.

Pricing has become another major battleground.

Alibaba offers Qwen3.8-Max at approximately:

• $2 per million input tokens
• $6 per million output tokens

In comparison, Kimi K3 is priced at:

• $3 per million input tokens
• $15 per million output tokens

These pricing differences could significantly reduce operating expenses for businesses deploying AI applications at scale.

Why Model Size Isn’t the Only Cost Factor

While parameter count often attracts headlines, it is only one factor influencing AI costs.

Several technical elements determine the actual expense of running an AI model, including:

• Model architecture
• Active parameter count
• Token usage
• Number of inference calls
• Output length
• Context window size

A larger model can sometimes operate more efficiently than a smaller one if its architecture minimizes unnecessary computation. Likewise, lower API pricing does not always translate into lower real-world costs if a model generates excessive output or requires repeated interactions to complete a task.

This is why many organizations now evaluate AI models using cost-per-task rather than token pricing alone.

Qwen3.8-Max Performs Strongly in Industry Rankings

Performance remains an important factor alongside pricing.

Following its release, Qwen3.8-Max reached the top position among Chinese language models on the crowdsourced comparison platform Arena.AI. The model also secured the second position on Arena.AI’s leaderboard for multimodal models capable of understanding images and visual content.

Although Alibaba’s model performed exceptionally well, several Anthropic Claude models continued to rank higher overall in global comparisons.

These rankings indicate that Chinese AI developers are rapidly closing the performance gap while offering increasingly competitive pricing.

DeepSeek Focuses on Affordable AI Inference

Rather than building the largest possible model, DeepSeek has chosen to compete primarily on affordability.

Its latest model, V4-Flash, features approximately 284 billion total parameters, with only 13 billion active parameters during inference. This lightweight activation significantly reduces computational costs while maintaining competitive performance.

DeepSeek also supports a one-million-token context window, allowing it to process extensive documents and long conversations efficiently.

The most striking feature, however, is its pricing.

According to Artificial Analysis, DeepSeek V4-Flash costs only:

• $0.14 per million input tokens
• $0.28 per million output tokens

These rates are substantially lower than many competing commercial AI models, making DeepSeek one of the most affordable options currently available.

Cached Inputs Reduce Costs Even Further

DeepSeek further lowers expenses through aggressive cached input pricing.

Artificial Analysis reports that the Max Effort version of V4-Flash charges only $0.003 per million cached input tokens, representing roughly a 98% reduction compared with its standard input pricing.

Cached inputs refer to previously processed information that can be reused during later requests without requiring the model to recompute the same context.

For enterprise applications involving repeated conversations, coding assistants, document analysis, and long-running workflows, caching can dramatically reduce infrastructure costs.

Benchmark Results Highlight Real-World Efficiency

Benchmark testing provides additional insight into model efficiency.

Reuters reported that Artificial Analysis estimated the average testing cost of DeepSeek V4-Flash at approximately $0.03 per benchmark, compared with:

• $0.86 for Kimi K3
• $1.86 for OpenAI GPT-5.6 Sol
• $3.15 for Anthropic Claude Fable 5

These results demonstrate that overall workload costs depend not only on API pricing but also on token consumption and computational efficiency throughout an entire task.

Artificial Analysis also awarded the Max Effort reasoning version of V4-Flash an Intelligence Index score of 40, while recording output speeds of approximately 118 tokens per second during testing.

Cost Per Task Matters More Than Token Pricing

Moonshot AI’s Kimi K3 illustrates why evaluating complete workloads is becoming increasingly important.

Although its API pricing remains competitive, Artificial Analysis found that Kimi K3 averaged approximately $10.57 per task on its AA-Briefcase benchmark. The model generated around 120,000 output tokens while requiring an average of 83 interactions to complete each task.

This example shows that higher output volumes and repeated model calls can significantly increase total operating costs, regardless of attractive token pricing.

Businesses are therefore shifting their focus toward overall workload efficiency instead of comparing only price per million tokens.

Open-Weight AI Models Expand Deployment Options

Another major trend shaping China’s AI ecosystem is the growing popularity of open-weight AI models.

Alibaba, DeepSeek, and Moonshot AI all continue supporting open-weight releases alongside hosted API services.

Artificial Analysis lists DeepSeek V4-Flash as an open-weight model released under the MIT License, with model weights available through Hugging Face. Similarly, Kimi K3 is distributed under Moonshot AI’s own licensing framework.

Open-weight models provide developers with greater flexibility by allowing deployment on private infrastructure or third-party cloud platforms rather than relying exclusively on vendor-hosted APIs.

Although infrastructure costs remain the responsibility of developers, this approach offers improved transparency, customization, and deployment freedom.

Businesses Want Affordable and Practical AI

Industry analysts believe many organizations no longer require the world’s most powerful AI model for every application.

According to Lian Jye Su, Chief Analyst at Omdia, businesses increasingly prioritize AI systems that are affordable, transparent, accessible, and capable of meeting practical business requirements rather than achieving the highest benchmark scores.

As AI adoption expands across industries, efficient open-weight models are expected to play a growing role in enterprise deployments.

Conclusion

China’s AI industry is rapidly shifting from a competition centered on model size to one focused on cost efficiency, practical performance, and deployment flexibility. Alibaba’s Qwen3.8-Max, DeepSeek’s V4-Flash, and Moonshot AI’s Kimi K3 demonstrate different strategies for balancing performance with affordability.

While Alibaba emphasizes powerful multimodal capabilities and competitive pricing, DeepSeek is redefining affordability through exceptionally low inference costs and efficient sparse architectures. Moonshot AI continues to compete with large-scale models and strong benchmark performance, highlighting that no single metric determines the best AI system.

For developers and businesses, choosing an AI model increasingly depends on total workload cost, deployment options, and real-world efficiency rather than model size alone. As the market evolves, lower-cost AI solutions are likely to accelerate enterprise adoption and shape the next generation of artificial intelligence across China and beyond.

Frequently Asked Questions

1. What is Alibaba Qwen3.8-Max?
Qwen3.8-Max is Alibaba’s largest AI model, featuring a Mixture-of-Experts architecture, multimodal capabilities, and a one-million-token context window.

2. Why is DeepSeek V4-Flash gaining attention?
DeepSeek V4-Flash offers extremely low inference pricing while maintaining competitive performance, making it an attractive option for developers and enterprises.

3. What are open-weight AI models?
Open-weight models allow developers to download and deploy model weights on their own infrastructure instead of relying solely on hosted API services.

4. Why is cost per task more important than token pricing?
Total workload cost depends on token usage, output volume, model interactions, and architecture—not just the advertised API price.

5. Which companies are leading China’s AI model competition?
Alibaba, DeepSeek, and Moonshot AI are among the leading Chinese companies competing through advanced AI models, lower inference costs, and flexible deployment options.