Anyone open to showing me a demo?

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10 Projects
1

I was wondering if anyone could show me a demo of the DGX Spark in action? I am considering getting a couple for us to do general LLM/AI work internally/locally here. I’d like to see:

  • General LLM chatting
  • Agentic coding capabilities (compared to Claude Code for example)
  • RAG, document searching examples

I’d also like to discuss what I can do with two Spark units and just confirm what I am trying to do is capable. I am very experienced with the commercial LLMs but would just like to confirm things before purchasing. Thank you.,

2

vast or other providers - like scan.co.uk will rent you units out so you can test - why would any community member give you a tour? call up your nvidia sales rep or rent the device - i dont want to be “this guy” but you asking for something that isnt even remotely in the scope or spirit of community

3

Sorry, I talked to NVidia sales and they told me to come here and ask.

4

What do you mean by a demo? It’s not clear to me if you’re asking for someone to post their experience here, do a video call with you, or meet you in person. I’m assuming @mrDragonFox assumed you’re asking for something in person, which yeah wouldn’t be likely to happen.

5

Yes, when I talked to Nvidia, I asked for a Zoom with them where they could share their screen of a dgx spark implementations and I could ask questions. Again, I apologize if this is not the right place.

6

There must be tons of youtube videos out there

7

I don’t think it matters so much that it’s not the right place, but if you want a screen-sharing session with someone I’d consider offering some form of compensation for their time, a gift card or whatever, otherwise it seems unlikely someone will step up for you.

Edit: I’m not even at the point of using RAG myself yet, and I’d bet a majority of users also aren’t.

8

OK thank you, I will reconsider my request.

9

Yes. I’ve run 1, 2, 4, 8, 16, 20, and 24. I just switched over to sparkrun to be more aligned with the community.

I have an exo Apple RDMA cluster attached.

10

Well I’m beyond interested in hearing what you’ve done with 16, 20, and 24 Sparks.

11

To get a quick start-
Check LLM benchmarks on Spark-arena.
Try opencode first.
Manually review

12

Thank you! I didn’t know about opencode.

13

Based on the questions you asked, I will say that DGX Sparks will likely disappoint you.

You have to first realize that it is an empty piece of hardware, and what it can do depends not just on its configuration, but very much on how you set it up, and therefore your technical background.

To get a feeling about what it can do, here are a few practical steps for you:

  1. Go through the posts on this forum to see what models you can serve on one to three DGX Sparks, this is the most common size of cluster on this forum. Pay attention to their quantization in addition to the model, as it can affect the quality of the model.
  2. Test the model out on Open Router to get a feeling about what these models can do for you, whether they are good enough for your use.
  3. Find out the context sizs, pp and decode speeds for these models (many test results on this forum), and see if the speed can serve your purpose. Be ready that you can only serve one to three users at most for most models with usable speeds.
  4. Use AI to help you go through the posts efficiently

Hope this helps

14

Thanks for the tips. Yes, I’m basically hoping I can use 2xSparks for one user to use as a coding agent, comparable to Claude Code. Maybe not 100% Claude Code quality, but close enough to be productive. I’ve read around here a little and other places, and it sounds like this is realistic, but I’ll check it out more. Maybe the company will let me rent the 2x for a month to test.

Also it seems as though the units are all out of stock.

15

@NVac https://spark.enverge.ai/ is renting DGX Sparks starting $0.75/hour. Give it a try!

16

Hi @NVac Not sure if my usage is very representative, but I’ve found I’m generally using my two sparks in conjunction with Frontier AI via Claude Code and Codex. So for me it’s not a question of local or frontier AI but a mix.

I completed a POC for a client which used docling and vision models for processing a UK LPA form with images for each pages and while doing this Claude Code was monitoring the work and evaluating/debugging this and then presenting the analysis and working through the options with me. So the inference used in the document processing was local with frontier AI orchestration as a an example.

I mostly use my two sparks independently but it is awesome to have the option to share the GPU’s for larger models if needed.

I’m running a software factory on my Dell GB10 with fine-tuned LORA adapters for the specific roles which I use vLLM to load the base model and LORA adapters which means I can host all six fine-tunes which with merged models there wasn’t enough memory. I then run Qwen 3.8 flash Next on the DGX Spark for the ‘Player’ coding agent. So I’m using the two sparks together for a single project solution but not sharing the GPU. I’m using a Qwen embedding model with Postgres and pg vector for the agent memory/job specific context - storing and retrieval of ADR’s, lessons learned etc.

I’ve got a Study Tutor working with ChromaDB Rag again I created the Langchain DeepAgents SDK agent and the RAG using Claude Code - it writes the agent and configures everything-I’m not doing any setup manually.

I’m planning on using the Pi coding agent harness with Qwen 3.8 flash Next in the next few days to evaluate it - in theory this gives me something similar to Opus from April/May this year which is pretty compelling given the recent usage limit cuts for the OpenAI ChatGPT Pro subscription.

Personally I can’t see the subscription subsidies lasting and want to have a means to build personal and client projects without spending thousands on Frontier AI API tokens. I’m trying to take advantage of the current subscription plans to build capabilities to future proof myself and learn valuable skills along the way.

I’ve actually never sat chatting with a locally hosted LLM. I use them with custom agent harnesses as part of a project.

I’ve used traces from execution on Opus to create a corpus from which to fine-tune local LORA adapters so again another example of mixing both local and frontier AI.

So before I brought the sparks I wanted to learn how to host models locally and fine-tune them and build custom agent harnesses. I thought that I would be able to use frontier AI less, but actually ended up using it more to build lots of things! So your current ideas for how to use them could change as you actually work with them.

I think they are amazing machines, but you have to understand the limitations upfront to manage your expectations and be realistic about what you can do with them.

Apologies for the essay but it’s something I find interesting to talk about and it isn’t discussed much in the forum - what people actually use these machines for!

17

Thank you, I appreciate the post. Yes, I understand your configuration and I would normally do it the same way. The only thing in this case is that we are not allowed to use the commercial LLMs, so I have to do all the AI with the Spark units. But from what I can tell, it seems I should be able to reasonably accomplish the things i want with the coding agentic functions, it’s just it might be slower and not totally frontier model level, which is fine I think. Even models from last year commercial are good enough.

18

if you are not allowed to use commercial llm’s you wont be allowed to use chinese llm’s either - what you can pick from is very “sparse” .. and given the restriction i assume its corporate or defense - better buy would be a b200/b300 box then as the spark is really targeted to 1-2 users tops who dont use it as productions mashines - its a r&d env for cx7 after all - it can be used for sure - but if you have regulatory constrains there should be the budget to actually buy the proper hardware you gonna thank your self later