Closed models have held their own against open-source models in 2025, but Chinese open-source models have eaten into the market-shares of all other open models.
According to OpenRouter’s State of AI report, proprietary models from major providers maintained a dominant 70% share of total token volume throughout the year, while open-source models collectively held steady at approximately 30%. However, beneath this stable top-line split lies a significant geographic shift: Chinese open-source models have substantially increased their footprint within the open-source category, displacing rest-of-world offerings.
The data reveals that Chinese open-source models now represent a growing portion of that 30% open-source share, with dramatic upticks following major releases. DeepSeek V3 and DeepSeek R1, along with Kimi K2 and the Qwen 3 Coder, triggered notable usage spikes that persisted well beyond their initial launch weeks. This sustained adoption suggests these models are being deployed in production environments rather than merely tested during brief experimentation phases.
The shift represents a meaningful rebalancing within the open-source ecosystem. While models like Meta’s Llama 3.3 70B continue to see usage, Chinese offerings have captured an increasing share of developers’ attention and token volume. The OpenRouter data shows that non-Chinese open-source models, which once dominated the open-weight category, now occupy a smaller slice of that 30% total.
This geographic redistribution has important implications for the global AI landscape. Chinese models are demonstrating that they can compete effectively on performance and adoption metrics, not just on price or accessibility. The sustained usage patterns following releases like DeepSeek V3 indicate that developers are finding genuine production value in these models, rather than treating them as experimental alternatives.
The 30% baseline for open-source models appears relatively stable across 2025, with the horizontal dashed line in OpenRouter’s visualization marking this OSS/proprietary split. Major releases from both Chinese and Western providers created temporary fluctuations, but the overall balance between closed and open approaches has remained remarkably consistent throughout the year.
For businesses evaluating their AI infrastructure, this data suggests two key takeaways: first, that proprietary models retain a commanding lead in overall usage, likely due to performance advantages in complex tasks; and second, that the open-source landscape is becoming increasingly influenced by Chinese providers, creating new options and competitive dynamics for developers seeking alternatives to closed models.
The persistence of usage spikes beyond initial release weeks is particularly noteworthy. Unlike consumer software, where new releases often see immediate adoption followed by decay, these open-source models appear to be finding lasting product-market fit. This pattern suggests that the quality gap between open and closed models may be narrowing in specific use cases, even as proprietary offerings maintain their overall market leadership.