Customer Ratings & Reviews

If you know what you're using it for, this is a great little box. Its real potential is better recognized when you have two, to compensate for larger models + context windows, but even with 1 you can do quite a bit. Works well for fine tuning, and can host models if you understand its limitations. Dense models are a bit slow on it due to limited memory bandwidth (around 20tk/s for something like Owen 3.8 27b), but it has extremely fast prompt processing. Makes it great for running automations that you don't have to interact live with. For things you do, use a MoE model like Qwen 3.6 35b a3b. I run that with Hermes on it, and it's quite fast for useable work. The FE model's 4tb storage is also very useful to have if you switch models for different purposes. And I think it looks the best out of all the GB10 devices. Just keep in mind airflow is not that great, and I just hook up a couple of usb fans in front of it to keep it cool.

Amazing machine. It runs Linux, which is a plus. You can run it with a mouse and keyboard. I run it headless, ssh into the box via NVIDIA Sync. You can run Ollama; I switched to vLLM. You can run lots of models from Hugging Face. It is slower than I running in the cloud, but your output is saved on your machine. It is smaller than I thought it would be. It does get warm. It is well designed, simple, nice color; the metal grill is very nice. I added a JCreate docking station. I added an external drive that was easy to mount. The 4TB drive has a Linux desktop; I disabled it to have the full 128GB drive for a model. With the drive, you can download different models to try them out. I should have gotten two. One DGX Spark is good for one model. With a second, you can run two or more smaller models. It is a little slower for token generation, but it does get the job done. It is quiet, and I sometimes couldn't tell it was running. The NVIDIA Sync has a nice resource monitor, which is very useful. When working on context, the GPU usage is 100%. It will get warm. I think they could sell a 256GB and a 512GB version. It would be nice if they used faster RAM. The speed is tolerable for me. With Hermes, you can do so much. Is it worth it? Yes. I would get a second one. Since I've gotten mine, the price has gone up. I run it wirelessly, and it works well. You can run VS Code on it. Linux is so nice on it. Learn the commands, and you have power. I'm using it to develop. I use it at home. Is it as good as a cloud LLM? It's good enough for me. You can use both. And with 128GB, it can hold all your code if you're building. This is the main thing I noticed: it seems quicker on the cloud. Maybe because it can see the whole repo. It is worth it.

I’ve been really impressed with the DGX Spark. It gives me a compact, quiet, and secure way to run serious AI workloads locally without jumping straight into a much more expensive workstation or server-class setup. All of that that is also easier portable due to it's very small size. What I’ve done with it has been exactly the kind of work I wanted from a system like this: local model experimentation, running larger LLMs that benefit more from memory than raw CPU, and testing AI workflows without depending entirely on cloud inference. The 128GB of unified memory and the DGX software stack make it especially useful for AI development, and it feels built for exactly that kind of hands-on work. It can do a lot of the cloud AI work but in no way shape or form replaces the online subscription cloud models. What makes it great is that it strikes a rare balance. It’s powerful enough to be genuinely useful for developers and researchers, but still far more approachable as an entry point than larger AI infrastructure options. Compared with machines in the DGX H100 class or other enterprise systems, this is a much lower-cost way to get into local AI hardware while still keeping strong performance and flexibility. For me, the biggest win is that it lowers the barrier to serious local AI work. Instead of needing a full rack, noisy cooling, or a cloud bill that keeps growing, I can prototype, test, and iterate on my own desk. If you want a practical AI machine that feels like a real step up without going all the way to enterprise pricing, the DGX Spark makes a lot of sense. I have now reasoning model with Gemma 4 and a Coder model, Qwen3 Coder. Completely load and ready to be used at any time. Or I can load one bigger model and do dynamic load and unload ah hoc. For my cons, I would say I wish that was even more memory and it would be more cheaper. But even at this price point, I think it is still fair.

I have been using the NVIDIA DGX Spark (GB10) primarily for running local LLMs, multi-agent pipeline development, and heavy quantitative backtesting. The performance packed into such a small deskside footprint is outstanding. 128GB Unified Memory: Having high-capacity shared memory allows running ~70B parameter models and large context windows locally without hitting standard VRAM limits or PCIe throughput bottlenecks. Quiet Form Factor: Even during continuous compute workloads, the thermal management keeps the unit running cool on the desk with minimal noise. Turnkey Software Stack: NVIDIA DGX OS comes preconfigured with CUDA and TensorRT-LLM, which made setting up local AI agent environments straightforward on day one. The only minor consideration is that the Arm-based CPU core layout occasionally requires compiling specific Python libraries or C++ extensions from source depending on your environment. If you need a reliable, high-performance machine for local inference, testing, and data-sensitive workflows without relying constantly on cloud clusters, this workstation delivers exceptional capability.

Excellent choice on AI computer. Very reliable. Pricy and fast delivery.

This is the way to build your own person ai server

This can run so many MoE models nicely. It won’t give you speed of cloud providers but a versatile work horse.

Best computer... IT's working like a charm running my models and agents.