As a LLM engineer, I would say what DeepSeek contributed is way more than the majority thinks. First, it helped many local LLM enthusiasts and many enterprises for their in-house usage, which was cool, and they definitely won the huge huge applauses in the past 15 months and been considered one of the best AI labs worldwide. But more importantly, if you look deeper into many of the open source datasets as well as the domain-specific open source models on huggingface, most of them are distilled from DeepSeek's model (and also Qwen, one of the GOAT lab for sure). And these datasets are mostly used by researchers or small companies who wish to do some researches or post train their own model, and you know what? They could be the next ones that change the world. If you think about this entire chain reaction, one open model offers people (or even the entire human being) so many possibilities, all the way from local user to industry to academia. They can save lots of money on investing computational power instead of distilling GPT for a one-time-use dataset, this i the real treasure that some of us often ignore.
I don't call myself an AI researcher (unfortunately) bc I am not smart at all, but I used to work as a LLM engineer and apply strategies on pre-training data, the model (or the models that based on our pretrained base model) is still now ranked top 10 in OpenRouter. We used tons of DeepSeek R1 internally to do data labelling, cleaning and all the dirty things. What I can tell is: these things 100% couldn't be done without a near-SOTA-and-open-weight model unless you would pay 10x cash from calling GPT/Claude's api. And I am pretty sure we are not the only lab who did this, and I even doubt that Claude/GPT/Gemini used DeepSeek-R1 at some point (I know it is very very unlikely but who knows).
We should appreciate what DeepSeek had done, as a human being. Basically they did not open source the model, but open sourced millions of possibilities and money to the entire AI world.