Not much had been heard from DeepSeek since it had stunned the world with its R1 model late last year, but the Chinese AI lab seems to be back in the game with a bang.
DeepSeek has released an update to its viral R1 model. The newer version is dubbed DeepSeek-R1-0528, and the company says it performs far better than its original R1 model. More impressively, the model seems to hold its own against OpenAI and Google’s top models.
DeepSeek-R1-0528 is here!” DeepSeek posted on X. The company added that the new model had improved benchmark performance, enhanced front-end capabilities, reduced hallucinations, and supported JSON output and function calling. Like its previous model, DeepSeek has released the weights of the new R1-0528 model. It is already available to try for users on its website.
But where DeepSeek-R1-0528 really seems to shine is on some benchmarks. On AIME 2024, it scores 91.4, just a shade under OpenAI o3’s 91.6, and ahead of Google’s Gemini-2.5-Pro-0506. On AIME 2025, DeepSeek-R1-0528 is a shade below OpenAI’s o3, and ahead of Google’s top model. On the GPQA Diamond, Aider, and Humanity’s last exam, DeepSeek-R1-0528 is just behind OpenAI and Google, while on LiveCodeBench, it beats Google, but is behind OpenAI. On all these benchmarks, DeepSeek-R1-0528 is comfortably ahead of its previous R1 model.
These are extremely impressive results, and show that DeepSeek R1 was no flash in the pan, and the company is still managing to come up with cutting edge models. In particular, DeepSeek R1-0528 beats Google-Gemini-Pro-2.5 on several benchmarks. Google’s top models are all confidential and proprietary, but a Chinese company hasn’t only created a model that beats Google on several benchmarks, but has open-sourced the model to boot. This is undoubtedly spur on competition — OpenAI has said that it intends to release an open-source model soon, and would likely want it to perform better on benchmarks than DeepSeek’s open-source model. Google too would like to make sure that its top models are state-of-the-art, and wouldn’t want to be bested by a Chinese open-source lab. And interestingly, DeepSeek hasn’t dubbed this release R2 — it’s simply called it an update to the R1 model — which could indicate that DeepSeek might have an even more powerful model in the offing. But the release of an open-source model that’s close to the state of the art proprietary models once again throws the AI race wide open, and shows how China is mounting a serious challenge against the top US AI labs.