Nvidia RTX Spark

I added an R9700 32GB to my 10+ year old desktop that had a 980 4GB card in it, for a grand total of $1350 or so. The payoff compared to what I was using with GHCP was 33 months, but when GHCP announced their price increase, it basically became a 3 month payoff at minimum (so yes, GHCP did a 10x price increase for non-parallel agentic workflows)

I can easily run Qwen3.6 35B-A3B with Q5_K_M with a 260k+ context window with some vram to spare. It easily runs probably 80tps. It took me quite a while to find the

Compared to GHCP Claude Sonnet 4.5 or 4.6, I have full parity. The wall clock time is faster for agentic workflows, and rule following is about on par.

With either, doing something kind of novel or obscure takes more hand holding compared to just generate a GUI or crud app. For example, trying to build an actual program that performs a complicated process correctly requires quite a bit of hand holding to get it to properly help.

Sure, it isn't Opus or something, but I think with the right harness, it probably can get close. I think most of the issues these days is the harnesses are lacking.

This plus the price different had me buy an AMD Strix Halo board last week. It seems the work with vLLM and training models could make the Spark worth the price difference, but before today's news I had the same thought about support and did not want to lock myself into a cool paperweight, especially with 128gb of RAM on the line. AMD is x86 and I can repurpose that or run Linux forever.

LM Studio Link is brilliant, outside their central login/auth requirement. Tailscale is the backbone, I think, so it makes sense but I'm sure a method with wireguard could exist and enable similar performance.

the current dielmma for me is how do I install a model on a remote LM Studio device without bypassing Lm Studio to SSH or remote in?

> lms link [servername] get model ?

> lms get [servername] model ?

> lms get model --link [servername] ?

Maybe I need to read the docs again but I swear the only way is remote or go to that device and download via the GUI, ssh in and use the local cli.

Maybe can copy/paste from one device's downloads dir to the server? Maybe I need to try hosting models on my NAS and see if I can download from device 1 then run on device 2 without install/setup?

If it's something like:

- v4.5: 1x cost, 100% quality, 100% speed but maybe sometimes 80% speed because of load - v4.6: 3x cost, 105% quality, 80% speed most of the time depends - v4.7: 9x cost, 115% quality, 90% speed most of the time

Then people will either stick with v4.5 for everything it can do and, if knowledgeable, use v4.7+ for critical or specific tasks.

But if we add the option of:

LocalLLM: one time hardware + electricity cost, good enough quality for 90% of work, good enough speed for 90% of work, no vendor lock in/sudden cost spikes...

Then there is an edge to running it yourself unless you can burn investor cash to get to the next level.

I think the recent headlines on org token spend plus my own experience just today (June 1) with the new Copilot Pro limits is going to push those with the compute to run locally.

As of about 1pm today I did something to hit 47% of my entire June premium requests (copilot Pro, not converted).

As of 2pm I'm using Gemma 4 E4B on a 12gb GPU (with large context window) off my desktop to power VS Code with Copilot on my laptop. I'm going to build an AMD Strix Halo system next week when parts arrive so I can queue up a few models in parallel or work with something I need that much RAM for.

I'm not lifting the earth with my LLM setup. Gemma 4 E4B is solid for accelerating my current projects. and it's costing me pennies more per hour vs blowing half my Copilot Pro plan in a distracted morning.

I'm at a vendor conference this weekend that is showing off their Agent/Agentic workflows. Nobody can tell me how they balance the cost long term. Hopefully whoever the vendor is paying for their cloud LLM token usage doesn't spike cost in a year (or the vendor themselves) after companies convert and are trapped VMware style with these agent processes. You can bring your own (cloud) model subscription. I need to find out if we can point it back to our own local LLM endpoint and try local models for the same processes. Even if it takes 5x longer, it could be cheaper and more secure.

I've never had a single problem with my Nvidia GPUs on Linux.