DeepSeek V3

The largest thing we serve, in six regions rather than thirty-one.

$argosy up --model deepseek-v3

ttft58ms

p50 from the nearest warm region

Throughput132tok/s

Single stream, no batching

Context128K

Full window, no sliding truncation

Price$0.58

Per million tokens served

The largest thing we serve. It needs eight cards per replica, so it lives in six regions rather than thirty-one, and requests from elsewhere are routed to the nearest of those.

Deploymentdep_c20ad4

Parameters671B MoE

Quantisationfp8

LicenceDeepSeek

Home regionsin1

Extra replicas2

p50 latency62ms

p99 latency148ms

Requests / s640

Figures are illustrative — this is a demonstration site.

Last 30 minutessin1

cache hitqueue 0ms errors 0

Also in the catalogue

LL Llama 3.3 70B Instruct 38ms $0.34 LL Llama 3.1 8B Instruct 19ms $0.06 QC Qwen2.5 Coder 32B 27ms $0.18 MX Mixtral 8x7B Instruct 31ms $0.22

Put the model next to the user.

One command, thirty-one regions, and an invoice that matches what you actually served.

Deploy a model Talk to an engineer