Ten leading Chinese developers introduced 16 new models in September, including tools for generating audio and video, Nikkei Asia reported.
Nikkei Asia counted models that were not merely repackaged or optimized versions of existing systems. Audio and video generation tools can be misused for impersonation.
BenchLM counted 20 large language models released by eight major Chinese developers from July through September. The same group released two models in the first quarter and 17 in the second. U.S. developers released 59 models from January through September, about 50% more than their Chinese counterparts.
Chinese developers are focusing on smaller, less expensive models, which are in demand in part because AI agents consume large volumes of tokens. Alibaba said in September that it planned a Qwen variant with 5 trillion to 10 trillion parameters. The previous flagship model in the series, released in August, has 2.4 trillion parameters.
Leading U.S. models remain faster, but experts say the gap is narrowing. 3DNews reported that Chinese models can cost 20 to 50 times less than Anthropic's systems for the same task.