DeepSeek open-sources the inference stack it uses to serve models at scale

Grace Parker · 13 likes

respect for actually shipping open source instead of 'open' marketing

Luciana Vargas · 6 likes

meanwhile some labs call a model card 'open'

Miguel Santos · 8 likes

Does it run on non-H-series hardware? thats the real question for most of us

Isha Menon · 10 likes

readme mentions H800 and H20 mostly, consumer gpus not a priority

Midnight Runner · 6 likes

their MLA + expert parallel tricks were always the secret sauce. having the code is huge

Pablo Martín · 13 likes

vllm already merged parts of their earlier kernels, this probably accelerates that

Emma Tremblay · 5 likes

i'll wait for someone to dockerize it properly lol

Urban Atlas · 6 likes

give it 48 hours

Isha Menon · 3 likes

the economics angle matters. if serving gets cheaper for everyone the API price war gets brutal

Isha Menon · 8 likes

good for us users tho

Elsa Andersson · 1 likes

MiniMax already cut prices this week, its happening

Priya Nair · 3 likes

license?

Bianca Santos · 8 likes

MIT for the serving code from what i saw, check before shipping obviously

Rapid Coder · 1 likes

another stack to evaluate 😮‍💨 my team still hasnt finished evaluating the last 3

Jack Taylor · 2 likes

evaluation fatigue is real

Hamza Malik · 0 likes

just wait 2 weeks, someone will post benchmarks on twitter, half of them wrong