Solved NVIDIA DGX Spark (GB10)

You still need an ARM64-compatible setup and a vLLM/model combination that supports the hardware and quantization you want.

What GPU would you use for a 27B dense model?

I would look at GB10 first if you want a simple single-machine setup with room for context and cache.

A 5090 can also be excellent if your chosen quantization fits inside 32 GB. That can give you much higher raw inference speed.

For Qwen3.8 27B specifically, AxForge already runs an NVFP4 version on GB10-class hardware. If you want dedicated hardware and simplicity, GB10 is the safe starting point; if speed matters more and the build fits, test a 5090.

What should I use for vision models with lots of images?

That depends on the vision model and how many images you process together.

If the workload fits in 32 GB, RTX 5090 is attractive because of its speed. If you need a large language model, vision encoder and large context resident together, GB10 gives you much more memory headroom.

For a production recommendation we would need the exact model, image sizes, images per request and target concurrency.

Is the RTX 5090 faster than the GB10?

Usually, yes, for a workload that fits comfortably on one 5090.

But that comparison gets misleading very quickly. A 5090 has much less memory, while GB10 is designed around fitting larger AI workloads into one 128 GB pool.

So I would not choose between them from a single tokens-per-second number. First check whether the model and context fit. If it fits in 32 GB, 5090 is often the speed-first choice. If it does not, GB10 may be the much simpler machine.