DeepSeek revolutionizes AI with new models - A competition for OpenAI!

DeepSeek revolutionizes AI with new models - A competition for OpenAI!

On May 29, 2025, Chinese startup DeepSeek released a new FOCUS for its R1 reasoning model. The R1-0528 model was presented on the developer platform Hugging Face, but without an official announcement. Details about the model itself, including how it compares to other AI models, have not yet been published. In the current LiveCodeBench rankings, DeepSeek's R1 inference model ranks third in code generation, just behind OpenAI's o4 mini and o3 models.

DeepSeek's R1 model is also reported to outperform its competitors, including xAI's Grok 3 mini and Alibaba's Qwen 3. A DeepSeek representative said in a WeChat group that a "small test upgrade" of the model is currently underway. Despite widespread belief that US export controls are hindering the progress of Chinese AI research, DeepSeek has emphasized that its newly released AI models are equal or even superior to those from the US, and at a lower cost.

Relevance to the tech industry

The launch of the R1 model in January also led to a decline in technology stocks outside China. Major Chinese companies like Alibaba and Tencent have since released 700 new models that are said to outperform DeepSeek. In this competitive landscape, Google has responded by introducing discounted access tiers to its Gemini model. OpenAI also lowered the prices for its services and introduced an o3 Mini model that requires less computing power.

DeepSeek is already planning to introduce its successor model, the R2, which was originally scheduled for release in May 2025. The company had already presented an upgrade for its large language model V3 in March of this year.

Innovative AI models

In addition to the R1 model, DeepSeek has introduced two new AI models: DeepSeek-R1-Zero and DeepSeek-R1. The R1-Zero model was trained exclusively using reinforcement learning (RL) without relying on supervised data. This shows that AI models can develop the ability to reason without human input.

The DeepSeek R1 model leverages initial data and achieves performance comparable to OpenAI's models in mathematics, programming and reasoning. DeepSeek has also released six more compact models that inherit the knowledge of the larger models through distillation. Among them, the 32B model outperforms OpenAI-o1-mini in various benchmarks.

All of DeepSeek's new models are open source and can be used via the DeepSeek platform or locally, making the technology even easier to access.

For more information please visit Reuters or City of Bremerhaven.