Solved NVIDIA DGX Spark (GB10)

The console is the place to check current availability. AxForge supports both immediate and scheduled bookings.

Can a GB10 run a 120B model?

Maybe, but the parameter count by itself is not enough to answer it.

A heavily quantized or MoE model can be realistic where a 120B dense model may be too large or too slow for the result you want.

AxForge's GB10 page shows large MoE models running on the hardware class, but for a specific 120B model we would want the exact model name and precision first.

If you send us the Hugging Face model link, we can size the weights and likely KV-cache requirement before recommending the machine.

Can I rent just one RTX 5090 instead of all four?

The current AxForge 5090 product is listed as a dedicated four-GPU system.

So I would not assume single-card 5090 rental is available from that product today. Register the requirement if that is what you need.

Can I run a different model on each of the four RTX 5090s?

Yes, that is one of the straightforward ways to use a four-card system.

Each 5090 has 32 GB of its own memory, so you can run separate model processes per card if each workload fits.