DeepSeek R1 open source model released; Competitors OpenAI o1 but 95% less expensive

Chinese AI startup DeepSeek has just released DeepSeek-R1, which is an open source and argumentative model. According to this startup, this model beats OpenAI’s o1 model in some benchmarks. It is noteworthy that the costs of the Chinese model are 95% lower than the American model.

While publishing the DeepSeek-R1 model in Hugging Face, DeepSeek has shown the performance of this reasoning model in various benchmarks in a report. R1 model has been able to beat o1 model in AIME, MATH-500 and SWE-bench Verified benchmarks. This model scored 79.8% in AIME math tests and 97.3% in MATH-500. It also scored 2,029 in the Codeforces test; It outperformed 96.3% of human programmers. On the other hand, the o1 model scored 79.2%, 96.4% and 96.6% in these benchmarks, respectively.

Of course, this model scored 90.8 percent in the MMLU benchmark, which is related to general knowledge, which is slightly lower than o1’s 91.8 percent. In general, the scores of these 2 models are very close to each other, with the difference that the Chinese startup model performs these calculations at a cost of 90-95% less than the o1 model.

Comparison of DeepSeek-R1 open source model with o1 model

Reasoning models are slower than normal models and usually take seconds to minutes to answer, but their advantage is that they are more reliable in areas such as physics, science, and mathematics. In its recent report, DeepSeek said that R1 contains 671 billion parameters and was developed based on the DeepSeek V3 model; Models with more parameters usually perform better than models with fewer parameters.

Of course, DeepSeek has also released smaller versions of R1 in the sizes of one billion parameters to 70 billion parameters. The smallest model can be run on laptops, but you need a more powerful system to run the full version of R1; Of course, if you are hesitating between using this model and the OpenAI model, it is better to take a look at the costs of these 2.

The o1 model costs approximately $15 per million input tokens and $60 per million output tokens, but the R1 model costs only $0.55 per million input tokens and $2.19 per million output tokens.

Of course, you can use the capabilities of this company’s reasoning models in the app and the free web version of DeepSik. To use its reasoning model, just activate the DeepThink option and state your request.