AVERI's Miles Brundage says deploying US open-weight models took hours compared to seamless Chinese alternatives
Florian Brand suggested OpenRouter to simplify the deployment process.
(cc @natolambert )
Miles Brundage@Miles_BrundageTFW you spend a few hours struggling to get American open weight models working on various clouds while Kimi and DeepSeek are plug and play
(earlier version of this tweet specified open source which really has been the main pain point, e.g. getting OLMo and Marin variants going, + should have kept that in in retrospect - there are American options but there tends to be a performance penalty and/or they're opaque)
Miles Brundage@Miles_BrundageTFW you spend a few hours struggling to get American open weight models working on various clouds while Kimi and DeepSeek are plug and play
#20Miles Brundage@MILES_BRUNDAGE@xeophon
Miles Brundage@Miles_Brundage(earlier version of this tweet specified open source which really has been the main pain point, e.g. getting OLMo and Marin variants going, + should have kept that in in retrospect - there are American options but there tends to be a performance penalty and/or they're opaque)
@Miles_Brundage check out our partner cirrascale's apis https://www.cirrascale.com/ai2endpoints
Miles Brundage@Miles_Brundage(earlier version of this tweet specified open source which really has been the main pain point, e.g. getting OLMo and Marin variants going, + should have kept that in in retrospect - there are American options but there tends to be a performance penalty and/or they're opaque)
@Miles_Brundage which one, openrouter has a ton of them?
Miles Brundage@Miles_BrundageTFW you spend a few hours struggling to get American open weight models working on various clouds while Kimi and DeepSeek are plug and play
@Miles_Brundage is it as simple as kimi? no. but also the more apt. comparison would be something like openbmb or internlm
Florian Brand@xeophon@Miles_Brundage hm, at least if you don't need max performance, renting a gpu and asking claude/codex to set up vllm and the correct parser etc is what i do and will give you an endpoint in like 10-20 minutes