True, but the newest mistral model is already pretty great

By the way: People have started questioning the numbers. Seems it’s not super clear whether Deepseek told the truth. And IMO the implications aren’t that clear either. If you can do the final training run of a singular model for $5 million… It still might require your parent company to build a datacenter for $1.6 billions and then rent the GPUs to you for $2 an hour. So it’s not like Europe can cough up a few millions and compete with OpenAI.

(German article here. But all of this is more speculation than anything else. At least as of now: https://www.heise.de/news/DeepSeek-V3-Entwicklung-soll-viel-teurer-gewesen-sein-10267750.html )