Chinese startup DeepSeek, which recently demonstrated its large-scale language model DeepSeek V3, has unveiled a new version of its AI, DeepSeek-R1. The developers claim that it is as good as OpenAI’s «thoughtful» o1 model in terms of performance and affordability.
Chinese startup DeepSeek, which recently demonstrated its large-scale language model DeepSeek V3, has unveiled a new version of its AI, DeepSeek-R1. The developers claim that it is as good as OpenAI’s «thoughtful» o1 model in terms of performance and affordability.
DeepSeek-R1, like o1, was trained using reinforcement learning (RL), but DeepSeek says it also applied supervised fine-tuning to handle complex reasoning tasks and match o1’s performance, VentureBeat reports .
To demonstrate the benefits of its approach, DeepSeek used R1 to distill six Llama and Qwen models, taking their performance to a new level. In one case, the distilled version of Qwen-1.5B outperformed much larger models, GPT-4o and Claude 3.5 Sonnet, on separate math tests.
These models, like the basic R1, were developed with open source and are available on Hugging Face under a license from the Massachusetts Institute of Technology.
During testing, DeepSeek-R1 scored 79,8% on the AIME 2024 math test and 97,3% on the MATH-500 test. It also scored 2,029 on Codeforces, beating 96,3% of human programmers. On these tests, the o1-1217 version scored 79,2%, 96,4%, and 96,6%, respectively. On the MMLU general knowledge test, R1 fell slightly behind, with an accuracy of 90,8% compared to o1's 91,8%.
The effectiveness of DeepSeek-R1 is being hailed as a major achievement for the Chinese startup in the AI space, which is currently dominated by companies from the United States. In addition, DeepSeek operates on an open source model and even provides access to educational materials.
Another advantage of DeepSeek for users is its pricing policy. OpenAI provides access to o1 at a price of $15 per million input tokens and $60 per million output tokens. In contrast, DeepSeek Reasoner, based on the R1 model, costs $0.55 per million input tokens and $2.19 per million output tokens.
The model can currently be tested on the DeepSee k chat platform, which resembles ChatGPT. Users can also access the model weights and code repository via Hugging Face, under the MIT license, or use the API for direct integration.
Recall that in DeepSeek’s internal benchmarking, the DeepSeek V3 model, on which R1 is based, outperforms both downloadable, «openly» available models and «closed» AI models that can only be accessed via APIs. In a series of coding competitions on the Codeforces platform, DeepSeek is ahead of other models, including Meta’s Llama 3.1 405B, OpenAI’s GPT-4o, and Alibaba’s Qwen 2.5 72B.
