Deepseek-R1-0528: Deepseek Update brings Chinese AI model back at eye level with western industry leaders

Published on: May 31, 2025 / update from: May 31, 2025 - Author: Konrad Wolfenstein

Deepseek-R1-0528: Deepseek Update brings Chinese AI model back at eye level with Western industry leaders-Image: Xpert.digital

Open Source AI at the limit: Deepseek performs Openai and Google in the shade

From 60 to 68: Deepseek catapulted Chinese AI back to the top

With the publication of deepseek-r1-0528 on May 28, 2025, the Chinese Ki Startup Deepseek achieved an important milestone that has redefined the global AI landscape. The update of the open source readering model shows dramatic performance increases and for the first time positions Deepseek at the same level as Openais O3 and Google Gemini 2.5 Pro. It is particularly noteworthy that this top performance is achieved with a fraction of the costs and with completely open model weights, which raises fundamental questions about the future of proprietary AI systems. The independent rating platform Artificial Analysis classifies the new model with 68 points - a jump from 60 to 68 points corresponds to the performance difference between Openaai O1 and O3.

Suitable for:

Deepseek and Alibaba: breakthrough at specialist level? The Chinese AI push in the healthcare system

The update and its technical improvements

Deepseek-R1-0528 represents a substantial further development, which does not require any changes to the basic architecture, but achieves significant increases in post-training through algorithmic optimizations and increased use of arithmetic resources. The update primarily focuses on improving the reasoning skills and, according to Deepseek, enables “significantly deeper thinking processes”. A particularly impressive example of this improvement shows in the Aime 2025 mathematics test, where the accuracy rose from 70 percent to 87.5 percent. At the same time, the average number of tokens per question increased from 12,000 to 23,000 tokens, which indicates more intensive processing processes.

In addition to the Reasoning improvements, the update introduces important new functionalities, including JSON output and functional views, an optimized user interface and reduced hallucinations. These innovations make the model much more practical for developers and significantly expand its scope. The availability remains unchanged: existing API users automatically receive the update, while the model weights are still available under the open co-license on Hugging Face.

Benchmark performance and performance comparisons

The benchmark results of Deepseek-R1-0528 show impressive improvements across all evaluation categories. In mathematical tasks, the Aime 2024 value rose from 79.8 to 91.4 percent, HMMT 2025 from 41.7 to 79.4 percent and CNMO 2024 from 78.8 to 86.9 percent. These results position the model as one of the most powerful AI systems for mathematical problem solutions worldwide.

With programming benchmarks, Deepseek-R1-0528 also shows significant progress. LiveCodebech improved from 63.5 to 73.3 percent, Aider polyglot from 53.3 to 71.6 percent and SWE verified from 49.2 to 57.6 percent. The Codeforces rating climbed from 1,530 to 1,930 points, which classifies the model in the top group of algorithmic problem solvers. Compared to competing models, Deepseek-R1 reaches 49.2 percent at SWE Verified and is therefore just ahead of Openaai O1-1217 with 48.9 percent, while Codeforces with 96.3 percentages and an ELO rating of 2029 points are very close to Openais.

General knowledge and logic tests confirm the broad increase in performance: GPQA-Diamond rose from 71.5 to 81.0 percent, Humanity's Last Exam from 8.5 to 17.7 percent, MMLU-Pro from 84.0 to 85.0 percent and MMLU-Redux from 92.9 to 93.4 percent. Only with Openais Simpleqa was a slight decline from 30.1 to 27.8 percent. These comprehensive improvements document that Deepseek-R1-0528 is not only competitive in specialized areas, but across the entire spectrum of cognitive tasks.

Technical architecture and innovations

The technical basis of deepseek-r1-0528 is based on a highly developed MOE (Mixture of Experts) architecture with 37 billion active parameters from a total of 671 billion parameters and a context length of 128,000 tokens. The model implements advanced purforcement learning in order to achieve self -check, multi -stage reflection and the ability to argue that is tailored to humans. This architecture enables the model to manage complex reasoning tasks through iterative thinking processes, which differentiates between traditional voice models.

A particularly innovative aspect is the development of a distilled variant, Deepseek-R1-0528-QWEN3-8B, which was created by distilling the thoughts of Deepseek-R1-0528 for the post-training of QWEN3-8B base. This smaller version achieves impressive services with significantly lower resource requirements and can be run on GPUS with 8-12 GB VRAM. The model achieved State-of-the-Art performance in the Aime 2024 test under open source models with a 10 percent improvement compared to QWEN3-8B and comparable performance such as QWen3-235B-Thinking.

The development methodology shows that Deepseek increasingly relies on post-training with Reinforcement Learning, which led to a 40 percent increase in token consumption in evaluation-from 71 to 99 million tokens. This indicates that the model generates longer and deeper answers without fundamental architectural changes.

Market position and competitive dynamics

Deepseek-R1-0528 establishes itself as a serious competitor to the leading proprietary models of western technology companies. According to Artificial Analysis, the model with 68 points is on the same level as Google's Gemini 2.5 Pro and in front of models such as Xais Grok 3 Mini, Metas Llama 4 Maverick and Nvidias Nemotron Ultra. In the code category, Deepseek-R1-0528 reaches a level just below O4-Mini and O3.

The publication of the update has had a significant impact on the global AI landscape. Already the original publication of deepseek-R1 in January 2025 led to a break-in of technology shares outside China and questioned the assumption that the scaling of AI required enormous computing power and investments. The response from the western competitors was quick: Google introduced discounted access tariffs for Gemini, while Openai lowered prices and introduced an O3 mini model that needed less computing power.

Interestingly, text-style analyzes from EQBench show that Deepseek-R1 is more oriented towards Google than on Openaai, which indicates that more synthetic gemini outputs may have been used in the development. This observation underlines the complex influences and technology transfer between the various AI developers.

Cost efficiency and availability

A decisive competitive advantage of Deepseek-R1-0528 is its extraordinary cost efficiency. The price structure is significantly cheaper than that of Openai: input tokens cost $ 0.14 per million tokens for cache hits and $ 0.55 at Cache Misses, while output tokens cost $ 2.19 per million tokens. In comparison, Openai O1 requires $ 15 for input tokens and $ 60 for output tokens per million, which makes Deepseek-R1 over 90-95 percent cheaper.

Microsoft Azure also offers Deepseek-R1 with competitive prices: The global version costs $ 0.00135 for input tokens and $ 0.0054 for output tokens per 1,000 tokens, while the regional version has slightly higher prices. This pricing makes the model particularly attractive for companies and developers who want to use high-quality AI functionalities without the high costs of proprietary solutions.

The availability as an open source model under co-license also enables commercial use and modification without license fees. Developers can operate the model locally or use various APIs, which offers flexibility and control over the implementation. For users with limited resources, the distilled 8 billion parameter version is available, which runs on consumer hardware with 24 GB memory.

Suitable for:

China's catch -up in artificial intelligence: the case Deepseek and the strategic data use

China's AI catching up: What the success of Deepseek means

Deepseek-R1-0528 marks a turning point in global AI development and demonstrates that Chinese companies can develop models despite US export restrictions that compete with the best western systems. The update proves that significant performance increases without fundamental architectural changes are possible if post-training optimizations and re-forcement learning are effectively used. The combination of top performance, drastically reduced costs and open source availability questions established business models in the AI industry.

The reactions of western competitors to Deepseek's success already show the first market changes: price cuts at Openaai and Google as well as the development of resource -saving models. With the expected publication of Deepseek-R2, which was originally planned for May 2025, this competitive pressure could further intensify. The success story of Deepseek-R1-0528 shows that innovation in the AI does not necessarily require massive investments and arithmetic resources, but can be achieved through clever algorithms and efficient development methods.

Suitable for: