Is the US-China AI capability gap still meaningful for actual production workloads?
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
Claude model activity is surfacing — worth tracking for capability changes, ecosystem impact, and availability.
A developer notes increasing reliance on Chinese AI models like DeepSeek, Qwen, and GLM for production workloads, with Claude as a fallback. They observed that Hy3, despite its smaller 295B total parameters (21B active per token), produced output quality comparable to larger models like DeepSeek's 671B or Kimi K3's 2.8 trillion for coding and API integration tasks. This experience suggests that the capability gap between US and Chinese AI models for practical applications might be narrowing, prompting a reevaluation of perceived differences.
I've been using Chinese models more and more this year. Started with DeepSeek for reasoning stuff, moved to Qwen for longer context work, tried GLM when it had that mini DeepSeek moment on OpenRouter. At this point the rotation is mostly Chinese models with Claude as the fallback for tricky creative tasks.
This week I finally got around to trying Hy3 and it kind of drove the point home. This is a model that activates 21B parameters per token out of a 295B total. It's tiny compared to DeepSeek's 671B or Kimi K3's 2.8 trillion. And yet for the coding and API integration work I threw at it, the output quality was closer to those models than it had any right to be. That's the part that's hard to ignore.
When DeepSeek alone is dominating OpenRouter usage, Qwen is leading Arena-Hard, and now even a small efficiency-focused model like Hy3 is hanging with them on real tasks…the "moat" around OpenAI and Anthropic just doesn't match what I'm seeing day to day. If this is what a 21B-active model can do in mid-2026, I genuinely don't know what the gap argument is even based on anymore.
However that’s just my feelings, I’m curious what everyone else is seeing in their own stacks.