Best path to learning modelling, Spark or 5090 or Mac?

Posted by yes_i_tried_google@reddit | LocalLLaMA | View on Reddit | 13 comments

I’m spiralling on the options and, combined with my natural habit of overanalysing everything, I can’t figure this out! So help please :)

For context, I’m a Director of Software Eng in banking trying to learn LLM and modelling techniques in my spare time. We use Copilot at work, and I’ve been hammering Claude Code at home to try and keep up until the bank brings things in.

My workflow up until now.

Claude Opus designs the hard stuff, cross-check with Ollama cloud models then gets one of the cloud models to build stuff. Works great, quality awesome, here I’m happy in principle but it’s burning £300/m on subs between Claude/Ollama - which I max out weekly but the quality of output is near perfect.

Now, I want to learn more about modelling, and tuning/quanting etc. in these areas, I’ve zero knowledge. If I could move some of my workflow locally and cut the Ollama subscription, even better.

Options I’m considering….

DGX Spark / GB10 — £3,800-4,000

Corsair Strix Halo — £2,550

5090 workstation — £4,900

Also Mac, but the memory shortage limits what we can/can’t order in UK, combined with a 3 month wait on what we can

Models I use in Ollama. Qwen3.5, glm5.1, kimi2.5, minimax2.7, Gemma4

I’m spinning on these options, but I THINK the Spark. Please could I get some views? It would be headless, and I’d carry on working with my baby MacBook M1 8GB / VPS.

[-]

MixtureOfAmateurs@reddit

The Spark is best for fune tuning and quanting, any training like workload with high batch numbers.

Strix halo is best for inference with high batch numbers, or large MoEs. I wouldn't use it for training.

5090 best for single user inference and training little neural nets. If you don't end up training, just running one model at a time that claude manages this will be fastest.

I would go the spark, maybe try training on a VPS like vast.ai to see if it's something you're into first. If not a used 3090 or two will be cheaper than a 5090 system.

yes_i_tried_google@reddit (OP)

Thanks! Hadn’t occurred to me to try a VPS for this

CalligrapherFar7833@reddit

You can start learning modeling by posing infront of a mirror you dont really need a spark for that ?

tmvr@reddit

I mean, you kinda need to have at least some spark, otherwise it's not going to go anywhere.

Lol

RainierPC@reddit

But can he turn left?

Actually, right now, no I can’t. My neck and back seized up 🫠 maybe im halfway there!

BankjaPrameth@reddit

What about multi-turn?

Im fat so no clue

magikfly@reddit

dude, if you're director of engineering why not use company resources to claim cloud compute? it'll be much more versatile than any of the options you mentioned.

Indeed…. Let me introduce you to my leadership.

vasimv@reddit

DGX spark would be good option for learning and experimenting (as with 128GB you can choose from many different models, quantizations, KV cache sizes, etc). And also will have option to run multiple models and programs at same time to experiment with powerful chat interfaces like open-webui with additional models to generate images, voice recognition, setup stuff like openclaw and so on.

If you want to do training fine-tuning, and quantization, you might be better off renting a cloud GPU