The Chinese artificial intelligence company DeepSeek has unveiled its new reasoning-driven AI model, DeepSeek-R1, on Monday. This launch marks the introduction of the complete version of the open-source model, following a preview release two months prior. The model is now accessible for download and can also function as a plug-and-play API.
DeepSeek-R1 AI Models Priced Up to 95 Percent Lower than OpenAI’s o1
The latest series includes two versions: DeepSeek-R1 and DeepSeek-R1-Zero. Both models have been derived from another large language model, DeepSeek V3, developed by the company. Utilizing a mixture-of-experts (MoE) architecture, the models merge multiple smaller components to enhance the overall efficiency and performance of the larger model.
Users can download the DeepSeek-R1 AI models from its Hugging Face listing. The model is licensed under MIT, permitting both academic and commercial use. For those who prefer not to operate the model locally, an alternative option is available through the model’s API here. The company has also published the inference pricing for the API, indicating that costs are 90-95 percent lower than those of OpenAI’s o1.
At present, the input price for the DeepSeek-R1 API is set at $0.14 (approximately Rs. 12.10) per million tokens, while the output price is $2.19 (around Rs. 189.50) per million tokens. In contrast, OpenAI’s o1 API charges $7.50 (approximately Rs. 649) for each million input tokens and $60 (about Rs. 5,190) for output tokens.
Aside from its lower costs, DeepSeek claims that its model delivers superior performance compared to OpenAI’s offering. According to the company’s internal testing, DeepSeek-R1 surpassed the o1 in several benchmarks, including the American Invitational Mathematics Examination (AIME), Math-500, and SWE-bench. However, the differences in performance were reported to be minimal.
In terms of post-training methodology, DeepSeek employed reinforcement learning (RL) on the base model without any supervised fine-tuning (SFT). This approach, often referred to as pure RL, grants the model greater flexibility in addressing intricate problems via the chain-of-thought (CoT) mechanism. DeepSeek has claimed that this project is pioneering within the open-source community, being the first to incorporate pure RL to enhance reasoning functions.
