Perhaps your prior comment would’ve been better received if it said that specifically instead of “Chinese models”.
But also, the latest DeepSeek is 1.6T parameters. “Choosing” to run this locally is a choice that comes with a seven digit price tag, and is a sunk cost that will probably not run any other frontier model anytime soon.
Most organizations are not looking to spend millions of dollars trying to find a workaround to specifically run DeepSeek. Most enterprise consumption in this space is still very experimental and a pay as you go model is much more palatable. Most are simply just looking for three checkboxes: is it close to frontier performance, is it compliant with my organizations requirements, and is it a good price? DeepSeek can only do two of the three at the same time.
> But also, the latest DeepSeek is 1.6T parameters. “Choosing” to run this locally is a choice that comes with a seven digit price tag
Unless you're specifically thinking about running the model at stock precision in a datacenter environment and generating ~100 tok/s or more on a 24/7 basis (the equivalent of a >$1000/mo spend even on the cheapest third-party APIs), that's very likely off by multiple orders of magnitude. Even then, experimentation can be done with cheap neoclouds on a pay-as-you-go basis.
The model we're discussing (Deepseek) is open weight.
Perhaps your prior comment would’ve been better received if it said that specifically instead of “Chinese models”.
But also, the latest DeepSeek is 1.6T parameters. “Choosing” to run this locally is a choice that comes with a seven digit price tag, and is a sunk cost that will probably not run any other frontier model anytime soon.
Most organizations are not looking to spend millions of dollars trying to find a workaround to specifically run DeepSeek. Most enterprise consumption in this space is still very experimental and a pay as you go model is much more palatable. Most are simply just looking for three checkboxes: is it close to frontier performance, is it compliant with my organizations requirements, and is it a good price? DeepSeek can only do two of the three at the same time.
> But also, the latest DeepSeek is 1.6T parameters. “Choosing” to run this locally is a choice that comes with a seven digit price tag
Unless you're specifically thinking about running the model at stock precision in a datacenter environment and generating ~100 tok/s or more on a 24/7 basis (the equivalent of a >$1000/mo spend even on the cheapest third-party APIs), that's very likely off by multiple orders of magnitude. Even then, experimentation can be done with cheap neoclouds on a pay-as-you-go basis.
5 replies →
My most sincere apologies for shortening "the vast majority of Chinese models" to simply "Chinese models".
I can see now why I was being downvoted - you have explained it eloquently.
(Your cost analysis is flawed and irrelevant. Azure serves V4 Pro.)