Open-weight models

Model Catalog

Open-weight models on a card nobody else shares, in an EU region you pick. Every build publishes its hardware, context length, license and hourly rate.

Available models

Every model runs on a GPU that is yours for the hours it is on — root SSH, no per-token fees, no rate limits.

Qwen3 0.6B,

0.6B

BF16
License Apache-2.0
BF16 €1.02/hr
GPU L4-1-24G
GPU VRAM 24 GB
Context 32,768 tokens
Throughput ~30 tok/s

qwen3.8

27B

FP8
License Apache-2.0
FP8 €1.02/hr
GPU L4-1-24G
GPU VRAM 24 GB
Context 262,144 tokens
Throughput ~70 tok/s

fp8 vs. full precision

fp8 stores each weight in 8 bits rather than 16, so an fp8 build holds the same model in about half the memory and often fits on a smaller, cheaper card. Full precision (BF16) is the build to pick when you would rather not trade any accuracy at all. We have not benchmarked the two against each other, so we publish no quality figure for the difference.

Not every model has an fp8 build. Each card above lists the builds we run for that model, with the GPU and the VRAM each one needs.

Ready to deploy?

Anyone can sign up. Our team turns on GPU access for each account, so tell us which of these you need to run.