0

I am not expert in the field, there seems to be a lot of discussion about definition of open source AI, I wonder what is DeepSeek in that terms, experts say that it is open weight, I've seen their Github and they have inference folder with python codes, also they have paper published, so I wonder is there some secret left about DeepSeek and how it achieved what it achieved? can everyone reproduce their chat app because of that openness, or is there some competitive secret?

math boy
  • 111
  • 3
  • 2
    I’m voting to close this question because I think this is not about AI (theory) but the meaning of open-source and it's specific to a tool. – nbro Feb 10 '25 at 22:43
  • There is not that much difference between the architecture of models. fine tunning is the key. So If u have hunders of A100 GPU, get the deep seek and fine tune it – M a m a D Feb 11 '25 at 08:34
  • 1
    @nbro I think definition of open source AI is currently important to the field of AI, open weight and open source are both terms connected to this field, so I think it is connected to AI theory, specifically about meanings of those "open source" in the context of DeepSeek. – math boy Feb 12 '25 at 21:18
  • Open-source is not specific to AI. It's related to software in general. – nbro Feb 12 '25 at 21:34
  • For software it has been clearly defined I think, however for AI, it isn't still clear what to call open source. Lex Fridman's last podcast discusses that topic well I think. – math boy Feb 13 '25 at 22:27

1 Answers1

1

As similar with other LLMs, DeepSeek provides model weights and inference code (for running the model) under the permissive MIT license, enabling users to deploy, fine-tune, and modify the models locally or via APIs. The code used to train the models and the datasets used for pre-training and fine-tuning remain undisclosed, a common industry practice due to cost and proprietary concerns. You can further refer here.

While some argue that without training code, a model cannot be considered fully open-source, others highlight that DeepSeek’s approach aligns with industry norms followed by leading AI companies like Meta, Google, and Alibaba.

DeepSeek is not an anomaly in how it approaches open-source AI. The practice of releasing model weights without training code has been the industry norm since Meta’s release of Llama 2. Companies like Google (Gemma), Alibaba (Qwen), and the GLM4 series have adopted similar policies. Even Llama 2 includes commercial restrictions, limiting usage for companies with over 700 million monthly active users.

cinch
  • 11,000
  • 3
  • 8
  • 17