This is a broad question to those of you who’ve reserved a DGX Spark…. What do you plan to use yours for?
For me, I’ve already built AI agents using existing cloud infrastructures and use AI as part of my daily workflow. I’m viewing the DGX Spark as an exploratory purchase. A way to better understand NVIDIA’s AI ecosystem and experiment with how their infrastructure connects from local compute to data center integration.
Moving to GB10 projects
In my case the reason is pretty straightforward. The last laptop upgrade cycle I was only focusing on increasing the configuration because of AI/ML workloads, but I’m still really satisfied with my M1 Max for everything else. The upgrade would cost me ~5k. At the same time, I still want to use my computer while the experiments are running, so it made sense to offload to another machine which is not in the cloud, doesn’t eat a lot of energy, doesn’t make a lotta noise and after it pays off, can be scaled-out or scaled-up. Once I got the budget approved, the Spark was a no-brainer. It has it’s trade-offs, but I’m ok with all of them and it’s a science lab you can carry around with not a lotta effort.
Prototyping AI solution for various use cases. We see a future with edge based AI devices connected to wearables/smart glasses and doing various vision based micro-services. The Spark is perfect for prototyping these.
I develop solutions that utilize AI - both LLM and more “traditional” ML workflows, and I can’t always use public API providers for privacy/compliance reasons, and running a development server in approved cloud (AWS/GCP/etc) is more expensive than buying a couple of Sparks. Ability to run a similar software stack as prod is also a plus.
Very cool. Not the reason I got the Spark but I’ve been toying with using wearables, primarily glasses with it.
I use the ASUS Ascent GX10 to validate hypotheses and build MVP projects based on clearly defined technical requirements. My goal is to quickly understand what a project is, how it should look, and to produce a tangible result without spending excessive time, effort, or money. Performance is not critical for me, so I simply provide the technical specification as input and receive a ready-to-use MVP project as output.
On a daily basis, it can incur costs of around $200 when calculated at GPT OSS 120B pricing, which means it pays for itself fairly quickly.
You get these?! ;)
Oh yeah, that’s a problem we’re very familiar with =)
It’s kind of like TDD, just in lazy mode 😄
You describe what should go in and what should come out.
How it’s implemented inside isn’t that important.
If the tests pass - well… it somehow works.
Sometimes, as implementers, we receive really solid, detailed documentation right away- which methods should exist, their types, attributes, what should go in and what should come out. In that case, it’s usually enough to just pass it to an LLM for implementation.
And sometimes the documentation isn’t sufficient - and then you basically have to write the entire spec yourself, down to the level of detail where either an LLM or a junior developer can clearly understand what is expected from them.
guess I am asking for too much…
Hey Guys,
what LLMs are you running?
I use vLLM with the model
Would also need OCR for handling PDF files and STT / TTS
Video generation would be sweet.
Kimi K2.5 would be sweet but it is way to big for a single Spark.
Guess there is not a single “omni” model with around 120B Parameters with a “OK Performance” that can do it all?
Dom,
I am planning on using Live Vlm playbook; where the system will analyze if an area is safe or not. Then send the information to my PLC to either allow a conveyor to run or forcing it to stop. My main pain right now is finding a way to send an input from the spark to my PLC. What are some findings from your investigation?
Martin,
Have you encountered any issues with the Qwen3 VLM?
its runs stable, is able to describe images, but not so “smart” as I have to say.. useless for openclaw so far. I do not know how to make the agent switch the model by itself. So for coding or text related content it should use a bigger model.
I guess that is what the machine is mainly being designed for.
still would love to have a personal local AI Assistant that is not sending data to “the cloud”.
Need a good text/code model and also Image and Video.
Martin,
Have you tried the Live VLM playbook?
Using it as a training tool for development practice.
Running Qwen3-Coder-Next-FP8 on the Spark via vLLM (~50 tok/s, 262K context) as the brain for OpenClaw — an agentic assistant reachable over Signal and WhatsApp. Mostly random experiments and slop so far.
Next up: Qwen2.5-VL-7B on a local RTX 3090 Ti as a vision model for browser automation — one model plans, the other sees the screen.
Thanks to everyone in this thread for the help and input — wouldn’t have gotten here without it.
@tsasse , the Live VLM Playbook is amazing and probably my favorite one so far. It’s not only unique in its ability to describe video in real time, but it’s also the easiest thing ever to get running; you don’t even need a virtual environment which I find makes it so much more convenient for future access. Just type “live-vlm-webui” into the terminal, including on my phone via Tailscale/Termius, and click on the URL. For me, it literally always starts with “a bald man…” which is not inaccurate, but literally every model I’ve used starts with the same reminder.
Yea right now I am waiting on a power supply for the camera I bought to test the system. The only pain point I am having is figuring out how to send a signal to a Turck node. I think I can connect the camera and node into a network switch but writing a script to send the information over is a little out of my wheel house.
On a side note do you run anything in tandem with the Live VLM Playbook?