While most of the world was still asleep Thursday, Chinese AI startup DeepSeek quietly dropped an upgrade to its R1 model—R1-0528.
It was released on developer hub Hugging Face.
No splashy press release, no grand claims.
Just a stealthy update that’s making some serious waves.
So, how good is it? According to the LiveCodeBench leaderboard—developed by AI heavyweights from UC Berkeley, MIT, and Cornell.
What’s The Innovation?
DeepSeek’s latest version now sits just behind OpenAI’s o4 mini.
It also trails the o3 models for code generation.
It is ahead of xAI’s Grok 3 mini and Alibaba’s Qwen 3.That’s no small feat.
A DeepSeek rep reportedly called it a “minor trial upgrade” in a WeChat group.
But the rest of the AI world is paying major attention.

Back in January, DeepSeek shook up the game when it launched the original R1.
It proved China didn’t need massive GPU farms or sky-high budgets to compete.
The move sent U.S. tech stocks wobbling and forced players like OpenAI and Google to rethink their pricing.
With R2 still looming on the horizon, one thing is clear: DeepSeek isn’t just playing catch-up—it’s sprinting.