Chinese AI startup DeepSeek is quickly gaining momentum in the global AI race. The company just released DeepSeek-R1-0528, proving once again that this is a bot to watch. The powerful update is already challenging rivals like OpenAI’s GPT-4o and Google’s Gemini.
The new version delivers major performance gains in complex reasoning, coding and logic, which are areas where even top-tier models often stumble.
With its open-source license and lightweight training demands, DeepSeek is proving to be faster and smarter.
A leap in benchmark performance
🚀 DeepSeek-R1-0528 is here!🔹 Improved benchmark performance🔹 Enhanced front-end capabilities🔹 Reduced hallucinations🔹 Supports JSON output & function calling✅ Try it now: https://t.co/IMbTch8Pii🔌 No change to API usage — docs here: https://t.co/Qf97ASptDD🔗… pic.twitter.com/kXCGFg9Z5LMay 29, 2025
In recent benchmark tests, DeepSeek-R1-0528 achieved an 87.5% accuracy on the AIME 2025 test.
This is a notable jump from the previous model’s 70%. It also improved significantly on the LiveCodeBench coding benchmark, moving from 63.5% to 73.3%, and more than doubled its performance on the notoriously difficult “Humanity’s Last Exam,” rising from 8.5% to 17.7%.
For those unfamiliar with what these benchmark tests mean, essentially, they suggest DeepSeek’s model can keep pace with, and in some cases outperform, its Western rivals in specific domains.
Open-source and easy to build on
Unlike OpenAI and Google, which tend to guard their best models behind APIs and paywalls, DeepSeek is keeping things open. R1-0528 is available under the MIT License, giving developers the freedom to use, modify, and deploy the model however they like.
The update also adds support for JSON outputs and function calling, making it easier to build apps and tools that plug directly into the model.
This open approach not only appeals to researchers and developers but also makes DeepSeek an increasingly attractive option for startups and companies seeking alternatives to closed platforms.
Trained smarter, not harder
One of the more impressive aspects of DeepSeek’s rise is how efficiently it’s building these models. According to the company, earlier versions were trained in just 55 days on roughly 2,000 GPUs at a cost of $5.58 million, just a fraction of what it typically costs to train models at this scale in the U.S.
This focus on resource-efficient training is a key differentiator, especially as the cost and carbon footprint of large language models continue to draw scrutiny.
What this means for the future of AI
DeepSeek’s latest release is a sign of shifting dynamics in the AI world. With strong reasoning abilities, transparent licensing, and a faster development cycle, DeepSeek is positioning itself as a serious competitor to industry heavyweights.
And as the global AI landscape becomes more multipolar, models like R1-0528 could play a major role in shaping not just what AI can do, but who gets to build it, control it and benefit from it.
More from Tom's Guide
- 7 genius AI prompts to supercharge your ideas (and get more done)
- I'm using Gemini now for my Gmail and there's one major discovery that's surprising
- I use ChatGPT to keep my houseplants happy — try these 5 game-changing prompts
You must confirm your public display name before commenting
Please logout and then login again, you will then be prompted to enter your display name.
