The DeepSeek folks just showed the world how to do the same thing those teams do, but at ~99% lower cost -- and published all code and weights as free open-source.
DeepSeek are great, however they didn't publish neither production code or their data pipelines. I still salute their openness in terms of architecture / great tech reports, but they keep their really performant training/inference code closed.
The DeepSeek folks just showed the world how to do the same thing those teams do, but at ~99% lower cost -- and published all code and weights as free open-source.