Content: Recently, China's DeepSeek team has launched its latest open-source large model R1, which has received widespread attention. The performance of the R1 model is exceptionally outstanding, surpassing OpenAI's o1 model in multiple tests, particularly excelling in assessments related to mathematics and programming.
In the latest AIME2024 test in the United States, R1 achieved a score of 79.8, surpassing o1's score of 79.2. In the MATH-500 test, R1 scored 97.3, also ahead of o1's 96.4. Additionally, in the SWE-bench Verified test, R1 scored 49.2, exceeding o1's 48.9. Although in the Codeforces coding test, R1 was only 0.3 points lower than o1, its overall performance is comparable to that of the o1 model.
Besides performance, R1's cost advantage is even more striking. OpenAI's o1 model charges up to $15 for every million tokens of input, while R1's cost is only $0.14, representing a 90% reduction. For output, o1 costs $60 per million tokens, while R1 only requires $2.19, achieving a reduction of 27 times. This significant cost difference makes R1 stand out in the field of open-source large models.
After the DeepSeek team announced the open-sourcing of R1, many foreign netizens expressed their admiration for this model, believing that R1 surpasses established open-source platforms like Meta and Mistral in terms of cost-effectiveness and performance. Many users stated that R1's efficient reasoning capabilities make it excel in code writing and mathematical explanations, with some even calling it "the model that resembles human inner monologue the most." Additionally, Awni Hannun, a machine learning researcher at Apple, tested R1 and found it to run quickly and efficiently on the Apple M2 Ultra.
The development of the R1 model went through multiple stages of training processes, including cold start data and multi-stage training, to enhance its reasoning capabilities and readability. These technical improvements ensure the R1 model's outstanding performance across various tasks.
With the release of R1, China's open-source large models have once again attracted significant attention and discussion in the international market, with many tech enthusiasts expressing their anticipation for the model's potential. The release of R1 marks a further breakthrough for China in the field of large model technology, promoting the development of open-source technology.
Open-source address: https://huggingface.co/deepseek-ai/R1
API: https://api-docs.deepseek.com/guides/reasoning_model
Key Points:
🌟 The R1 model outperformed OpenAI's o1 in multiple tests, demonstrating exceptional performance.
💰 The input and output costs of R1 are as low as $0.14 and $2.19 respectively, with a cost reduction of 90%.
🚀 R1 has received widespread attention after its open-source release, with many foreign experts praising its performance and high cost-effectiveness.
