- DeepSeek R2 promises up to 97% lower processing costs than GPT-4o.
- The model uses 1.200 billion parameters and has been trained on 5,2 petabytes of data.
- 82% of the chips used in your training are Huawei Ascend, reducing dependence on NVIDIA.
- Its efficiency and price could make it a serious global competitor.
The artificial intelligence sector is experiencing a true technological race, with major companies vying to develop more powerful and efficient models. DeepSeek, a Chinese firm, has burst onto the scene with its upcoming release: DeepSeek R2 , which, according to leaked data, could be a game-changer by offering high-capacity AI at significantly lower costs than its direct competitors.
The news has spread like wildfire in the tech sector, as DeepSeek R2 aims to become a viable and affordable alternative to giants like ChatGPT , Google Gemini , and Meta Llama. The desire to compete head-to-head is based on a compelling argument: the cost of processing large volumes of information will, according to estimates, be significantly lower than that of current leading models.
DeepSeek R2 Technical Features
DeepSeek R2 is presented as an open-source, generative artificial intelligence model, allowing both companies and individual developers to research, customize, and adapt the technology to their needs. Its new version implements 1,2 trillion parameters, raising the bar compared to its predecessor and approaching the performance of the most advanced options on the market.
The model was trained using 5,2 petabytes of data, mostly from the C-Eval 2.0 dataset, a detail that highlights the scale of the work undertaken. In terms of capabilities, its performance in computer vision is expected to achieve outstanding results, reaching up to 92,4% accuracy in tests with the COCO system.
The cost of use is one of its most important advantages. According to leaks, operating with DeepSeek R2 will cost just $0,07 per million tokens in and $0,27 per million tokens out, representing a cost reduction of nearly 97% compared to GPT-4o.
Technological independence with Huawei Ascend chips
One of the most significant changes in the development of DeepSeek R2 is its focus on domestically produced hardware . 82% of the chips used during training are from the Huawei Ascend 910B series, marking a clear departure from the use of NVIDIA GPUs, which has been the norm in the industry until now. This decision not only results in a drastic reduction in energy and production costs, but also greater technological independence, especially for the Chinese market.
This shift toward homegrown solutions allows DeepSeek to continue to move forward independently and reduce its dependence on the US supply chain, which could signal a shift in future AI development in both China and other emerging markets.
A global competitor against AI titans
In a landscape of constant innovation, DeepSeek R2 is emerging as a potential competitor to popular models like the GPT-4 Turbo or Google Gemini 2.0 Pro, thanks to its excellent price-performance ratio. Its focus on efficiency and scalability, along with its open-source philosophy, makes it appealing to both businesses and independent developers seeking more affordable alternatives without sacrificing performance.
For now, most of the available details come from leaks and analysis by AI experts, and we'll have to wait for official confirmation from the company. However, the excitement generated by this new model is undeniable and marks a turning point in the cost trend for generative artificial intelligence.
The trend toward greater competition in the field of AI continues, driving advancements that democratize access to these technologies. DeepSeek R2 represents a significant step in the search for more accessible and sustainable solutions, capable of democratizing access to advanced artificial intelligence both within and outside of Asia.
Passionate writer about the world of bytes and technology in general. I love sharing my knowledge through writing, and that's what I'll do on this blog, show you all the most interesting things about gadgets, software, hardware, tech trends, and more. My goal is to help you navigate the digital world in a simple and entertaining way.