At midnight, the well-known open-source large language model platform DeepSeek released its latest version R1 (0528), which has attracted widespread attention. This release did not come with detailed explanations; instead, DeepSeek chose to quietly release the model, with expectations that a model card would be published later to further introduce its features.
According to reports, the performance of the R1-0528 version was tested on the renowned code testing platform Live CodeBench, showing results comparable to OpenAI's latest o3 model. Many netizens expressed their anticipation and conducted tests on the new version, finding that the updated R1 performed similarly to OpenAI's o3 model in terms of style and performance.
In particular, R1-0528 demonstrated characteristics similar to o3-2.5 in professional-style responses. Whether in the way it uses arrows and asterisks or in the concluding "why it works" statement, R1 exhibited a high level of persuasiveness. Although OpenAI continues to enhance its model performance, DeepSeek’s R1 still performs impressively, sparking anticipation for its upcoming R2 version.
Hyperbolic Labs' co-founder and CEO also noted that R1 remains a model capable of precisely answering whether "9.9" is greater than "9.11," showcasing its high level of intelligence. Moreover, renowned AI commentator Haider had nothing but praise for the new R1 version, mentioning that during a programming challenge, R1 successfully built a word scoring system, providing clean code and executable test cases that passed perfectly on the first run.
Haider pointed out that previously only the o3 model could perform such tasks, but now DeepSeek’s R1 has reached this level. He mentioned that although the model's thinking time seems to have extended, its intelligence is impressive. As testing progresses, people can't help but wonder if DeepSeek is intentionally using this R1 version as a transition to prevent disappointment when R2 is released.
The new DeepSeek R1 performs exceptionally well, not only able to engage in self-correction like the o3 model through chains of reasoning but also possessing creative construction abilities akin to Claude, which were absent in previous versions. With the official release of the model card, the AIGC open community will provide deeper analysis, looking forward to more surprises in the future.
Open-source address: https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/tree/main
Key points:
🌟 The new DeepSeek R1 (0528) is open source, performing comparably to OpenAI's o3 model.
💻 Testing shows R1 excels in programming capabilities, successfully completing complex tasks.
🔍 Look forward to DeepSeek publishing a model card to deeply analyze the features of the new version.