We have tested our local DeepSeek installation. Since DeepSeek is an open-source LLM and large language model of the 2nd country in AI: China , is relatively easy to install on most computers, although its advanced configuration requires technical knowledge to access the code. The installation allows for offline operation to external servers, as shown in the video at the end of this report. Most users have observed that the model consumes more memory and processing power than expected, although its operation is smooth, as evidenced in our video.
Compared to other models, the performance of DeepSeek it is inferior to closed models such as ChatGPT version 4, according to Techcrunch and our own observations in Proportione. We have also detected bias and censorship, especially on issues related to Chinese politics, which has been corroborated by Wired and Ars Technica. Additional tests conducted by the University of Pennsylvania and Cisco in cybersecurity indicate that the model does not effectively block malicious tampering attempts.
In our tests in Proportione , when requesting a simple one-sentence summary, the model generated creative interpretations instead of selecting key information, which was curious but did not comply with the request. In an exercise in creativity, when asked to describe time without numbers or clocks, the model offered a lengthy and metaphorical answer, although not particularly concise. In Spanish, its expression is deficient, while in English it is adequate.
When asked about the inventor of radio, the model mixed answers in English and Spanish without coherence, mentioning only Marconi, although the final answer was acceptable. Their reasoning mode is detailed, which slows down interaction, but provides a thorough explanation of their process. A test that we have not recorded in the video was when we asked him about the recently deceased Spanish biochemist Margarita Salas, and he told us that she was a former president of Argentina who recently died as well, when we told him that this was not true he again put himself in the box that if it was true, we have to take into account that we are running it in the place and that his training is very limited, but things like these lead us to think that it is light years away from ChatGPT, another thing is that being open source it immediately improves very quickly. In summary, DeepSeek is promising and stands out for its unique and detailed approach to reasoning.
DeepSeek is moving fast, but is it enough to change the market?
In our analysis of DeepSeek Locally, we saw that although their model is open source, its actual performance still has challenges in terms of accuracy, resource consumption, and biases. However, one of the great advantages of the Open Source It's their ability to expand rapidly, and that's exactly what we're seeing.
In less than 48 hours, Azure, AWS, Snowflake, and Databricks have announced the integration of DeepSeek-R1 into their ecosystems. This move makes it easier for thousands of companies to deploy and test the model with a few clicks, without the need for large configurations.
What does this mean for businesses?
- Fast adoption, but no performance guarantees: Just because a model is available on large platforms does not mean that it is the best. It simply means that it is now more accessible to try it out.
- Open source accelerates dissemination, but not quality: OpenAI has shut down its model for strategic and commercial reasons. DeepSeek, being open, allows many companies to adopt it, but that doesn't solve its technical limitations.
- The market has yet to give its verdict: Availability on platforms like AWS or Azure doesn't mean companies will use it in large-scale production. It is a market test, not a definitive validation.
These types of integrations allow experimentation with DeepSeek, but it is still too early to know if it can really compete with more advanced models. Open Source accelerates expansion, but the question remains: Is DeepSeek good enough to change the business landscape, or is it simply riding the wave of open source to gain visibility?
At Proportione, we continue to test it and will analyze its evolution in real environments.