Comment by specked-citrus

> companies that might have reservations about chinese models having training backdoor are getting tools to train its own

This release is an inference stack, not training.

> the tricks deepseek uses to make building them faster can now be used by more companies, increasing competition even further.

But again, this is just inference, and specific to DeepSeek models. DeepSeek models are already dirt cheap on OpenRouter at pricing / speeds comparable to or better than DeepSeek's first party endpoint which uses this stack.

Yes, this repo will help others optimize inference, but it's nowhere near the hype being proclaimed in this thread.