DeepSeek upgrades R1, says performance nears OpenAI, Google models

The Chinese artificial intelligence company DeepSeek (DEEPSEEK), which roiled the tech world when it released its R1 large language model in January, released an upgraded version, which it claims is performing near the level of OpenAI’s o3 and Google’s (NASDAQ:GOOG)(NASDAQ:GOOGL) Gemini 2.5 Pro.

The new model is dubbed DeepSeek-R1-0528.

“In the latest update, DeepSeek R1 has significantly improved its depth of reasoning and inference capabilities by leveraging increased computational resources and introducing algorithmic optimization mechanisms during post-training,” DeepSeek said in a post today on Hugging Face. “The model has demonstrated outstanding performance across various benchmark evaluations, including mathematics, programming, and general logic. Its overall performance is now approaching that of leading models, such as O3 and Gemini 2.5 Pro.”

DeepSeek said the upgraded model’s accuracy has improved to 87.5% from 70% when handling complex reasoning tasks. It also said it has a reduced hallucination rate and is now better for vibe coding.

Last month, DeepSeek, released the open-sourced Prover-V2 model, a new specialist AI model. The company’s announcement came on the heels of Alibaba (BABA) unveiling Qwen 3, an enhanced AI model featuring hybrid reasoning capabilities.

In March, Chinese internet search giant Baidu (BIDU) launched its foundation model Ernie 4.5 and the deep-thinking reasoning model Ernie X1. Baidu had said then that X1 delivers performance on par with DeepSeek R1 at only half the price.