Alibaba Unveils Biggest Qwen Model; DeepSeek Drives AI Costs Lower

Alibaba Group unveiled its largest and most capable artificial intelligence model to date on Monday (August 3), while DeepSeek’s latest system offered running costs about one-hundredth of those of Anthropic’s flagship model in benchmark tests.

These developments show Chinese technology companies intensifying competition with US AI developers on two fronts. Alibaba’s Qwen3.8-Max is designed to challenge leading models on capability, while DeepSeek’s V4-Flash is seeking wider adoption through sharply lower prices.

Alibaba’s Hong Kong-listed shares rose as much as 7% after Qwen3.8-Max climbed global model rankings. The system has 2.4 Tn parameters, making it the largest model in the Qwen family and placing it close to Chinese rival Moonshot AI’s 2.8 Tn-parameter Kimi K3, launched last month.

Parameters are the numerical settings a model learns from data and uses to recognise patterns, generate responses and perform tasks. A higher parameter count does not automatically make a model more capable, but it indicates the scale of its architecture and has become a closely watched industry measure.

Qwen3.8-Max became the highest-ranked Chinese text model on crowdsourced comparison platform Arena.AI after its unveiling. It ranked behind Anthropic’s Claude Fable 5 and three Claude Opus variants.

The model placed second globally on a separate Arena.AI ranking, assessing its analysis of images and other visual material, and trailed only a variant of Claude Fable 5.

Qwen3.8-Max uses a mixture-of-experts architecture, which divides work among specialised parts of the system instead of activating the entire model for every request.

Only 95 Bn of its 2.4 Tn parameters are used at a time, helping reduce computing costs and response delays.

Alibaba said the model completed a software engineering project in 16 days. The company plans to release its model weights next week, allowing developers to download and adapt the system.

DeepSeek’s V4-Flash, meanwhile, is priced at $0.14 per Mn input tokens and $0.28 per Mn output tokens. Input already stored in its cache costs $0.0028 per Mn tokens, according to the company’s API documentation.

Artificial Analysis estimated that V4-Flash cost an average of 3 cents to complete each test in its benchmark suite. That compares with 86 cents for Moonshot’s Kimi K3, $1.86 for OpenAI’s GPT-5.6 Sol and $3.15 for Anthropic’s Claude Fable 5.

The per-test calculation provides a broader measure of cost than token prices alone because it accounts for how much data a model must process and generate to complete a task. A system with a low advertised price can still be expensive if it requires more steps or produces substantially longer answers.

DeepSeek rose to international prominence in early 2025 after its R1 and V3 models showed that competitive AI systems could be developed and operated at substantially lower cost than many leading US products. Their release triggered a sell-off in global technology shares and raised questions about the scale of American spending on AI infrastructure.

Alibaba and DeepSeek are both pursuing open-weight models, making their learned numerical settings available for developers to download and modify. Leading systems from OpenAI, Anthropic and Google generally remain proprietary, with access provided through applications and software interfaces.

“Many business workflows do not need the industry’s very best model,” said Lian Jye Su, chief analyst at research firm Omdia. “They need models that are good enough, affordable, transparent and accessible, and open-weight models help meet that demand.”

These launches reinforce a widening distinction in the global AI contest. US developers continue to lead many top capability rankings, but Chinese companies are narrowing the gap while exerting greater pressure on pricing, access and the economics of deploying AI at scale.