Looking at RTX Spark laptops for my employees for local agents. Wondering if anyone else is exploring the same, would love to chat.
Arm windows preference indicates some exquisite taste :) I am sure employees will be exilirated to spend half time battling compatibility issues and trying to have emulator work close to native.
Unless you plan to gut windows and go Linux only. Jenson kinda confirmed the ueifa is not locked. But you would be blessing micros lop with a license cost donation anyways. Imo either x86 win if you need windows stack or Mac if you need power and Ai in a laptop format
I honestly don’t understand obsession with having llm in a laptop. If you have employees and a company, rent instance at lambda, runpod, etc setup the model you want, have it automated, run 8-12 hours a day according to your schedules, with easy script to launch for some unscheduled workload. you can make it under 1000 usd a month a much greater performance and capabilities with full control over it than letting employees fiddle with own models, turning knobs and having fun instead of doing actual work. just imo
If you consider 128gb solution - you clearly not intending to run mid-tier models, just small ones. So having something like GH200 96 GB VRAM + 400 GB RAM for 2 USD per hour will give you massive concurrency and speed than any local solutions on individual notebooks. Set up vpn like tailscape for each and you are golden
If you want to own hardware and self-host → DGX Station solves your problem
For rtx spark to be viable some really really good 100-120b Moe model needs to be released.
And on the same topic: would you buy MacBooks pro m5 max 128gb for them right now?
Jenson will say but you have Nemotron 120B Super - just use it, lol :D
PS: if someone does not know this model is pathetic
MacBook Pro 16 128gb ram gets us further than any non-datacenter solution from Nvidia, specifically as end-user setup.
Support of gb10/sm12x from Nvidia nowhere near datacenter’s gpu, same for vendors (hp, dell, etc).
Apple is just apple, things just works and in very efficient way, plus employees genuinely happier with overall experience.
Those who got gb10 struggled alot, it is enthusiast solution, don’t expect 9-5pm employees to keep rebuilding sglang, vllm trt-llm just to get something to work, which simply works on MLX (llama.cpp, lm-studio, etc).
True if a single spark is considered vs single macbook/studio. No argument. Except cost → spark still much cheaper but it quickly changes :(
The difference is when you compare 2 sparks vs macs. You are limited to 128 gb ram and models that can be running there. 2 sparks are also faster. I don’t know how is it in US but in Europe Apple stopped even giving options to order anything with larger RAM than 128G. MacStudio is only offered in up to 96G. MacBook M5 Pro Max - 128GB. That is it folks, the walls are closing in.
But I reiterate - for a company of more than 2-3 people both options are suboptimal - go for a rented metal and run your own model or go big and get DGX Station for 100k and which will make sense if you need to run 24x7
RTX Spark laptops does not have Connectx-X7, and Macs supports now clustering with RDMA over TB5. You could connect 2-3 MacBooks or 4 studios.
For small companies inference as a service, personal subscriptions for claude, codex, etc, easier and cheaper than procurement/balance sheet/managing any high end hardware.
I have no idea who the RTX Spark is for, I am talking about DGX Spark. And you are completely wrong on cost here. My own calculations show 95% cheaper inference on 2 sparks for DeepSeek v4 Flash versus API cost of GPT5.4 mini, which gives comparable results for my tasks
sorry, was replying to topic started, author specifically mentioned RTX Spark laptop.
gotcha. The author provided almost no details. Maybe the need Word and Excel typers. But then why bother with local AI? use copilot goddamit! LOL
@nosajjason, it really depends on what your use case is, but I can’t imagine the RTX series being useful for what I’m imaging you’re talking about. First, you’d definitely want the maxed out ones with 128GB of RAM and not the cheaper ones. Second, with the RTX you’ll have no ability to cluster if you decide to scale that up. Both of these points are all the more true if you’re talking about agents for coding versus something that’s just written language like documents or reports.
We’re running multiple 8x DGX Spark clusters for agentic coding, which can run the current best-in-class open-source models. We’re only ever getting more and more happy with them over time. I couldn’t imagine trying to use a single DGX Spark for serious professional coding though. It’d really have to be a situation where the code for whatever reason can’t ever hit the cloud providers AND the business can’t afford to scale it up at all.
People can come from different directions, but in my view the Sparks are really only commercially relevant with at least a 4x cluster, but 8x really stands out in the marketplace. A single Spark, or two, would be great for a hobbyist that wants to experiment and learn while running something at home. That’s the same category the RTX devices will be in, except they’ll also have lower versions that will increasingly make less sense the lower they go. Just do some research on all of the Copilot+ PCs and people’s opinions on those to see where the lower-end RTX laptops will fall. LLMs really want as much memory as possible, even with all of the open-source memory efficiency improvements that are constantly rolling out.
I wouldn’t do RTX spark laptops. The main reason is because the models that can run on just 128gb are pretty limited in terms of intelligence. Sure qwen3.6 27b is good but lacks a lot of world knowledge due to only having 27b parameters and the 122b and 120b models are just not very good at anything unless its simple routine tasks IMO and even then they screw up simple web searches. Plus RTX spark cant be clustered with connectx7 so you cant connect them together to run large intelligent models. And from what ive seen in the marketing RTX spark mainly seems centered around gaming with AI as the second use case. If your wanting large model capacity and are investing a lot I think a DGX station would be a much better move than the equivalent in DGX sparks since at some point TP on sparks hits diminishing returns and because the DGX station has 288gb of 6TB/s GPU memory and ~396gb/s RAM and a total capacity of 768gb. Would be a much better experience IMO. If I was in the position to I would 100% get a DGX station it is a GB300 afterall.
I’d only go RTX if it was Linux and the 128GB version. It’d be nice for small to maybe medium sized local data science/scientific tasks and running small MoE models locally. I’ve done similar things with large RAM on a system76 laptop and SIMD.
But like others have said, your money is probably better spent on shared DGX clusters or runpod GPU so you can host the bigger OSS models.