Alibaba and DeepSeek Drive China’s AI Model Race with Cost-Effective Solutions

Alibaba has recently introduced its largest AI model, Qwen3.8-Max, with a whopping 2.4 trillion parameters. This model utilizes a mixture-of-experts architecture, activating only a portion of the model for each specific request. By doing so, around 95 billion parameters are active at any given time, reducing costs and response times compared to activating the entire model.

On the other hand, DeepSeek has also adopted a sparse architecture similar to Alibaba’s at a smaller scale. Their latest model, V4-Flash, boasts 284 billion total parameters, with 13 billion active during inference. In comparison, Moonshot AI’s Kimi K3 features 2.8 trillion total parameters, with approximately 104 billion active.

Qwen3.8-Max is versatile, capable of processing text, images, and videos while supporting up to one million tokens of context. Alibaba revealed that the model successfully completed a software engineering project within 16 days.

In terms of pricing, Qwen3.8-Max competes closely with Kimi K3, priced at $2 per million input tokens and $6 per million output tokens, significantly lower than Kimi K3’s rates of $3 and $15, respectively.

However, the cost of running a model is not solely determined by its size. Factors such as architecture, active parameter count, token consumption, and the number of calls needed to complete a task also influence the overall pricing.

Following its release, Qwen3.8-Max quickly rose to the top position among Chinese text models on the Arena.AI platform, although it still lagged behind several Anthropic models in the overall rankings. Additionally, it secured the second spot on Arena.AI’s leaderboard for models analyzing visual content, trailing behind an Anthropic Claude Fable 5 variant.

On the other hand, DeepSeek’s V4-Flash model stands out for its lower inference pricing strategy. Priced below several widely-used AI systems, V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens, as per Artificial Analysis.

Moreover, DeepSeek’s cache-hit pricing for the Max Effort version of V4-Flash is significantly lower at $0.003 per million tokens, allowing for substantial savings. In benchmark testing, V4-Flash demonstrated a cost of three cents per test, far lower than competing models like Kimi K3, OpenAI’s GPT-5.6 Sol, and Anthropic’s Claude Fable 5.

It’s important to note that the comparison of token prices provides only a partial view of the overall cost story. Factors such as output volume, number of model interactions, and repeated model calls can significantly impact the total cost of completing a workload.

Moonshot AI’s Kimi K3 exemplifies how advertised API prices may differ from the actual cost of completing extended workloads. While priced at $3 per million input tokens and $15 per million output tokens, Kimi K3 averaged $10.57 per task on the AA-Briefcase benchmark, showcasing the importance of considering various factors beyond token pricing.

In conclusion, the landscape of AI model deployment is evolving, with Chinese developers offering open-weight releases alongside hosted API access. This approach provides developers with more flexibility in how they deploy models, allowing them to run models on their infrastructure or through third-party providers.

In a nutshell, the cost of running AI models is influenced by a multitude of factors beyond token prices. By considering architecture, active parameters, token consumption, and deployment options, businesses can optimize their AI operations for efficiency and cost-effectiveness.