DeepSeek releases updated R1 reasoning model with significant improvements

Earlier this year, DeepSeek created a significant ripple in the tech industry. Despite being a largely unknown Chinese startup, DeepSeek released R1, a free and open-source reasoning model that performed on par with leading models from major AI companies, including OpenAI, Anthropic, and Google.

In addition to being open source and high-performing, DeepSeek demonstrated remarkable efficiency by using far fewer resources for training. This was achieved through several techniques, including a refined mixture-of-experts (MoE) architecture, advanced compression methods, multi-token prediction capabilities, and a reinforcement learning–first training methodology.

Today, DeepSeek released its first major update to the R1 model, DeepSeek-R1-0528. The updated model and its weights are already available on Hugging Face. Although DeepSeek has not yet published official benchmarks, several AI enthusiasts and developers have already begun sharing results — and they appear to be impressive.

According to LiveCodeBench, a benchmark for evaluating coding capabilities of large language models, DeepSeek-R1-0528 is now ranked No.4, surpassing models like o4 Mini (Low), Grok-3-mini (High), and Gemini 2.5 Flash Preview.

deepseek is so back...

the new version of deepSeek r1 is just incredible at coding

i run a coding challenge where the model builds a word scoring system

the model reasoned briefly and then gave me two files: clean code and working tests

both ran perfectly on the first try.… pic.twitter.com/HYztXPJ6Jb

— Haider. (@slow_developer) May 28, 2025

Additionally, X user @Chetasula reported that this latest DeepSeek R1 update now reasons deeply, similar to Google's models, and can "think" for up to 30-60 minutes depending on the task.

R1-0528 feels... aware 👀

🚨 New DeepSeek R1-0528 Update Highlights:
• 🧠 now reasons deeply like Google models
• ✍️ Improved writing tasks – more natural, better formatted
• 🔄 Distinct reasoning style – not just fast, but thoughtful
• ⏱️ Long thinking sessions – up to… pic.twitter.com/mkeQVIWlQk

— Chetaslua (@chetaslua) May 28, 2025

For those interested in trying out this latest R1 update, it is available on OpenRouter or LMArena.