Same 128GB but cheaper
Video Overview & Insights
Jetson Thor is $1,000 cheaper than DGX Spark with the same 128GB—so I benchmark it against Spark and an M4 Pro Mac mini to see what actually matters.
I would build a agentic business team to manage my t-shirt brands and also shift my usage from claude co-work to a local llm.
Try out ChatLLM - http://chatllm.abacus.ai/ltf and Abacus AI DeepAgent - http://deepagent.abacus.ai/ltf
My USB-C portable hub: https://amzn.to/4kw0hrf
Damn, those prices are crazy 😭😭
👀 My favorite external drive (dependable): https://amzn.to/3Os9Wi3
👀 Thunderbolt 4 dock: https://amzn.to/3yVRicC
From a cost-per-GB, or even cost-per-token-generated perspective, the Mac mini technically looks like the best value. What I’m not sure about is whether adding more RAM to the Mac mini, or switching to a Mac Studio, would make the value proposition comparable to the other setups.
👀 Thor on NVIDIA Marketplace: https://bit.ly/44j0acY
👀 Spark and Thor on Amazon: https://bit.ly/4pCjBpJ
6 months in and the spark is now $1,200 more and out of stock
👀 M4 Pro Mac Mini deal: https://amzn.to/3Mw8dNx
⚡ *Other gear I use:* https://www.amazon.com/shop/alexziskind
Ministral is highly capable. You probably should have work into your prompt engineering.
🎥 Related Videos 🎥
🧬🐍 Mac Studio CLUSTER vs M3 Ultra 🤯 - https://youtu.be/d8yS-2OyJhw
"What would you build with a Jetson Orin Nano Super?" would create my own offline personal assistant and do a whole lot of automation with different LLM's.
🧳🧰 Mini PC portable setup - https://youtu.be/4RYmsrarOSw
🍎💻 Dev setup on Mac - https://youtu.be/KiKUN4i1SeU
NVidia Jetson Thor is not so cheap nowdays, and especially in Europe
💸🧠 Cheap mini runs a 70B LLM 🤯 - https://youtu.be/xyKEQjUzfAk
🧪🔥 RAM torture test on Mac - https://youtu.be/l3zIwPgan7M
this would help me work on my projects where i create RL environaments to test out different LLM capabilities, i currently use hugging face pro for fine tuning and serving models, would love to have my own local machine
🍏⚡ FREE Local LLMs on Apple Silicon | FAST! - https://youtu.be/bp2eev21Qfo
🧠📉 REALITY vs Apple’s Memory Claims | vs RTX4090m - https://youtu.be/fdvzQAWXU7A
I would build and train image models for my ai influencer business
⚡💥 Thunderbolt 5 BREAKS Apple’s Upcharge - https://youtu.be/nHqrvxcRc7o
🧠🚀 INSANE Machine Learning on Neural Engine - https://youtu.be/Y2FOUg_jo7k
I would build an expression analyzer.
🧱🖥️ Mac Mini Cluster - https://youtu.be/GBR6pHZ68Ho
* 🛠️ Developer productivity Playlist - https://www.youtube.com/playlist?list=PLPwbI_iIX3aQCRdFGM7j4TY_7STfv2aXX
great benchmarking - been looking at getting some of these dedicated devices for llms - been using a thread ripper with 5090 currently as my home lab with llama and been looking to pick up a device to get bigger models to load. Great video.
— — — — — — — — —
❤️ SUBSCRIBE TO MY YOUTUBE CHANNEL 📺
Apple is winning the race and they are not completely involved in ai race yet
Click here to subscribe: https://www.youtube.com/@AZisk?sub_confirmation=1
Join this channel to get access to perks:
Consider running MLX on the Mac and see how much better it does.
https://www.youtube.com/channel/UCajiMK_CY9icRhLepS8_3ug/join
— — — — — — — — —
what I liked the most is keyboard design specifically picked and matching the machine desing )) everything needs to look perfect
📱LET'S CONNECT ON SOCIAL MEDIA
ALEX ON TWITTER: https://twitter.com/digitalix
NVIDIA bragging about RAM-rich computers is less impressive when we're not allowed to buy any. 128GB is half the price of their machine. So impossible for anyone to compete. Can't stop the corporate takeover of this space but I'm not going to cheer them on while they do it.
— — — — — — — — —
Raffle: To enter, subscribe to the channel and leave a comment on this video answering: “What would you build with a Jetson Orin Nano Super?” The prize is 1× Jetson Orin Nano Super signed by Jensen Huang. Entries are open from publish until 12:00 PM ET on Tuesday, December 30, 2025. One entry per person, 18+, open worldwide where permitted (void where prohibited). I’ll pick one winner at random, reply to their comment and post the winner in the YouTube Community tab, and they’ll need to respond within 72 hours to claim (any import duties/taxes are the winner’s responsibility). No purchase necessary; not affiliated with or endorsed by YouTube.
well this is somewhat unfair you should redo this with the M5 Max if you havent already
#dgxspark #nvidia #llm
Broh...those keyboards...giving me 1980 IBM vibes...peak satisfaction on every click?
More User Perspectives
Fix your Youtube award.
@pinkeye00Oooo shiny things.
@learnwhisperingVery cool benchmarking you do. I love it!
Would be nice if you could show us the values in tables... lot of numbers.. a bit confusing
Real benchmark video to understand.
Can you try running via MLX that optimized for Apple silicon ?
When I try to run Qwen3 coder in M5 Pro - 48 GB 14 inch laptop , I see the fan or temperature goes up to 95c while computing . is that expected ?
To get better comparison you should be using constant seed for each device
@adderek3:02 it’s called “excel”
@JonValtandtheEvilRobotsThe one thing DGX Spark gives you is the ability to use the QSFP112 DAC (Direct Attach Copper) 200G cable. This allows linking two NVIDIA DGX Spark units together & yields a combined memory pool of 256GB and up to 8TB of total storage capacity for distributed AI. Not sure Thor allows that? So in some ways, the extra cost is to allow easier scale?
@trigasurgeI would put all my effort into creating a real Jarvis-like-System. Hermes, Local voice, home assistant, Especially the Voice seems to be still not good. I Like the natural voice of ChatGPT and wanna recreate it for learning. In the end replacing the Alexas in my home with w real working local system would be great.
Additionally i would like to add this AI-guy to my work as i am doin remote support and it could help me a lot, if the AI just provides its experience.
153 what?
@Feherlof1aSame prompt on my i9 285k, 96gb ddr5 and RTX5080 16GB VRAM gave 18.05 tok/sec via gpt-oss 120b
@Daniel-kj8kji would build a tool that would create podcasts of the latest updates, releases and news in the ai sphere. I cant read everything but i do have time to listen so i can do other stuff.
@mennolaan2468I'd create a fully local chatbot and also host models for coding and homeassistant
@malvinportnerI would build a real-time object detection system using a Jetson Orin Nano Super
@stilllobster黄皮衣要是吃苹果的饭，库克一定会把滚烫的dgx spark 按在他脸上😂
@wqy8169I'm curious about the scalability of the MiniMax 2.7. Do you think it’s feasible to chain 4 of them together to work as a single node .
@gimmedogfoodHow many 100s of thousands is the channel omg
@t803586so they r for robots
@GeoffWong1983Thats a great giveaway, If i get to be the lucky one, I would, or I am building a web solution where artist get 1v1 feedback from other artist or an AI Artist that has the knowledge of the art style they want to draw inspiration from.
@peerviewi would build a new agent harness which i have in mind, inspired some by hermes agent, it'll feature self learning
the agent harness works on a different paradigm from other agent harness which work by the user typing a command or description in, and the ai does it
here, we instead have a sorta 'commandments', which the ai repeatedly, 24/7 works on, fullfilling, improving and fixing to comply, by never ending work
so you might write what you want, then go to sleep, and in the morning you might have 3-4 different compliant solutions for it
Did you run it by using MLX feature on Mac mini ?
@hacd10I recently started thinking about Local AI, and trying trying to figure out where to start. So I am watching you channel closely.
@PhysengineerI am an amateur researcher who uses machine learning to analyze stormwater patterns to establish more robust techniques.
@michaelz37304:15 "may not be as efficient" - I would argue that using memory adaptively could be a sign of more efficiency!
@JPeachyDevJETson is for Jenson Huang
@peacemercy7276In llama.cpp I had the exact problem the Thor has in 13:58 with my amd rx 7800 xt, it always went from 0% to 100% multiple times during pp, I changed from vulkan top rocm and It went from a pp of 130tps to up to 1500tps with 100% flat line gpu usage. Maybe the Thor needs some tweaking to avoid requesting the cpu for something (it's likely invoking cpu sometimes for unknown reasons, as vulkan was doing in my local)
@AlguienEnSuCasaSo if you have the money get the Nvidia DGX Spark, otherwise get the Nvidia Jetson Thor, but If im going to spend that much money I might as well go with the DGX Spark.
@kushtra24I'd go Spark for the x86 architecture.
@valdimer11Mac mini takes advantage of shared memory if you use mlx so ollama doesn’t make it shine
@antoinemanson4698Well...if it's just a hobby, you can tinker with this stuff. If you are a serious user the next step is a DGX Station..
@m3mee20105 token per second is practically useless for agentic work flow
@niteshyadav0903I think you should have benchmarkes a mlx formatted model on the mac as its a different architecture. Plus you then would need to make sure you used a llm runner that support mlx (lm studio, omlx etc)
@FurryButBaldI would update my openclaw to use a local model by default so then i can keep more of my queries local to me :)
@PaulHoughtonIndustrial inspection station
A fixed camera over a conveyor belt running segmentation models to detect defects, measure dimensions, or read codes. The Super variant's extra TOPS matter here when you need sub-100ms inference for high-speed lines, plus you can log analytics locally before batching to the cloud.