China's Open-Source AI Models Challenge US Dominance, Igniting Fierce Debate Over Performance Gap

The ongoing technological rivalry between the United States and China has found a new battleground in artificial intelligence, with the recent open-sourcing of Xiaomi Mimo 2.5 intensifying the debate over the performance gap between the two nations’ AI models. Xiaomi Mimo 2.5, now available with an MIT license, permits commercial use, fine-tuning, and further training without authorization. The suite includes a ‘Pro’ version, which reportedly ranks first among open-source models in various benchmarks for agentic tasks and code generation, and a standard omnimodal version (vision, audio, image) praised for its efficiency. This development has sparked contentious discussions: while some observers believe Chinese models are rapidly catching up to their US counterparts, others contend the quality difference is actually widening, with Chinese models reportedly trailing by as much as eight months, a claim supported by certain benchmark analyses like DeepSeek V4’s perceived GPT-5 level.

Contrasting views, however, suggest a much smaller gap, possibly only two to three months, with Chinese models steadily improving. This perspective is championed by figures like DAX from Anomal Co/OpenCode, who points to multi-benchmark analyses that paint a more favorable picture for Chinese AI. Criticisms against Chinese models often cite slower response times, though some argue this is frequently due to suboptimal native API usage rather than inherent model limitations, noting better performance when run on US-based servers. Another contentious point revolves around ‘distillation,’ with some alleging that Chinese open-source models achieve their performance by learning from proprietary, high-cost models. Despite these debates, a consensus is emerging among practitioners: regardless of whether the gap is two, four, or eight months, current open-source models, such as Mimo 2.5 (which some practitioners compare to Sonnet), are highly productive for a significant majority of professional tasks. This suggests that even with a lag, the rapid evolution of open-source AI means top-tier capabilities, akin to Opus 4.7, could soon become widely accessible without proprietary restrictions.