Look at this NVIDIA GX Spark! (vs Studio)

[rezwits]

NVIDIA DGX SPARK

I am looking at getting a MacStudio in the future to handle the likes of gpt-oss-120b-mlx-6Bit (96GB) etc models for private LLM or even larger models give or take in the future.

Now don't get me wrong, I understand that the DGX Spark would be way more optimized for AI and ML training if needed, but looking at the prices, it's interesting that:

A) The Spark is at 128GB unified (interesting huh?)
B) The Studio is at 96GB unified (each for the $3999)

That's sick that what small AI computing NPU Boxes are coming to.

I've read articles about multi-gpu boxes etc, but I have also read articles about, and have experience (a little), of GPU's dying in 6 months or 2-3 years on average.

I hate swapping GPUs etc

But this segment is interesting because it's like OK, 96-128 GB of RAM? $4000!

Now they have a higher costing model (not yet out) called the :

DGX Station (Blackwell Ultra / GB300 class) → The desktop big-memory beast.

• Up to ~784 GB coherent memory (≈288 GB HBM3e GPU + 496 GB LPDDR5X CPU), NVLink-C2C up to 900 GB/s, ConnectX-8 up to 800 Gb/s.

But look at that memory...

I mean for a 512GB Unified, from Apple you are talking about $9500, but the Nvidia? Probably the same if not more, in price.

Now you guys can put all your little haha's (fanboy etc) you want cause I think Apple isn't looking to shabby. I only say this because of Unified Memory.
But also of the fact of this consumer level product. I mean you get macOS and you can use the damn thing for other stuff.

So from a consumer point of view that's pretty nice. (I mean I like my 40-120 VRR HDR! just as much as the next guy! hehe)
But I mean if you look what segments we have, it's: Gaming Rig <-> Development Rig <-> AnIMaL Rig.
Which we know yeah if we could have gaming on mac sure, but I mean gees, come on with the games we have iOS, I still would rather get a PS6!

But this will put Apple right smack dab in the MIDDLE, which I think is a PERFECT PLACE to BE!

I am running a Mac Mini M4 Pro with 64GB and getting really good results from a Qwen3-Next-80B-A3B-Instruct-MLX-4bit (45GB) in Xcode it's pretty nice:

Private, Free, works in Xcode! Get's a little crazy sometimes if not put on a leash, but still!

But man a MacStudio M5 Ultra? for say $4000 with 128GB? or even bump up to 256GB Unified? for $5600 ~ $6000?

That's not bad from Apple IMHO...

I mean to not be a 100% AI company and to have a "side project" Mac Studio that is right there?! in the mix of all this? AMAZING...

source: MEDIUM ARTICLE

2¢

 

[maflynn]

What's you point? You post is all over the place, I'm not sure what you're asking/ranting/stating

The DGX Spark is a specific use type of computer deigned for AI. The Mac Studio is a general purpose computer that many of us use for many different tasks. Two different pieces of hardware for two different markets/customers

 

[rezwits]

There's no point, I am not trying to prove anything! I am just giving notice about the DGX Spark, so people can see they have a similar MacStudio option if they want to go the NVIDIA route for AI or if they are considering the DGX Spark, it's 100% possible to go the MacStudio route...

How is that not obvious that I am plugging information about the DGX Spark? This is a apple/mac site, I mean?

Laters...

 

[JSRinUK]

The Spark 128GB vs M4 Max Studio 128GB. For about the same price.

Pick your platform.

People buying Mac know they’re paying the Apple-tax for the upgrade (I know I did when I went all in on 128GB). Are people who don’t buy Mac prepared to pay that premium? Or would they rather spend that money on a dedicated Nvideo GPU? I suspect the latter.

The Spark just feels like they are playing catch-up to something Apple released 8 months ago.

Now, if the Spark had come out with 256GB RAM for ~£3,000 - £3,500? Then that would be something.

As it is, if you need 128GB RAM you now have your choice of platform. Take your pick. The other side isn’t going to tempt you away from that choice.

 

[maflynn]

The Spark just feels like they are playing catch-up to something Apple released 8 months ago.

Why do you say that?

 

[JSRinUK]

Why do you say that?

Because Apple released a “128GB unified memory AI machine” 8 months ago. The Spark should have jumped that with 256GB, not matched it.

 

[Basic75]

The Spark just feels like they are playing catch-up to something Apple released 8 months ago.

The Spark also has 200Gbps networking which is so far ahead of anything Apple offers.

 

[maflynn]

Because Apple released a “128GB unified memory AI machine” 8 months ago. The Spark should have jumped that with 256GB, not matched it.

The Studio? That's not an AI machine, but a general use computer that can also do AI - provided you spec it out sufficiently. Where as the spark is an AI machine - it does one thing and one thing only, and as such it doesn't have to contend with the overhead of a full blown operating system.

I'm not saying that the Studio is ill fitting for AI tasks, or buying the studio instead of the Spark is a poor decision.

As for the Spark, So far, I've only seen positive remarks, reviews, and YTs tbh, I don't recall anyone saying that nvidia is playing catchup, quite the opposite they're leading the charge in AI and this is but one piece in the puzzle

 

[JSRinUK]

The Studio? That's not an AI machine, but a general use computer that can also do AI - provided you spec it out sufficiently. Where as the spark is an AI machine - it does one thing and one thing only, and as such it doesn't have to contend with the overhead of a full blown operating system.

I'm not saying that the Studio is ill fitting for AI tasks, or buying the studio instead of the Spark is a poor decision.

As for the Spark, So far, I've only seen positive remarks, reviews, and YTs tbh, I don't recall anyone saying that nvidia is playing catchup, quite the opposite they're leading the charge in AI and this is but one piece in the puzzle

Most YouTube reviews are hyperactively positive but some people (who buy it with their own money) are a little more grounded in their opinions.

If I was into buying “an AI machine”, I wouldn’t be buying the Spark. It’s too compromised to get it into such a small package. Having a set-up with dedicated Nvidea GPUs is the path to go down for that.

You’re right that the Studio isn’t a “dedicated AI machine”, but if you can only spend one $4,000 and you want a machine that can stand its ground on AI as well as everything else you need it to do, you don’t have many other choices.

I don’t see a compelling reason to buy the Spark, unless you’re dead-set against Apple but you still want 128GB unified memory to play with. Or, if you want to buy multiple Sparks and connect them together (which I believe is a feature) - in which case, you’d probably still be better off spending that kind of money on dedicated Nvidea GPUs.

It’s difficult to know what gap in the market has been identified that requires the Spark to fill it.

 

[maflynn]

If I was into buying “an AI machine”,

I've not yet gotten into AI, and local LLMs, and so I can't say if its good or bad. I've played with them a little bit on my studio, but it became quite clear I need more ram then the base Studio came with

 

[JSRinUK]

I've not yet gotten into AI, and local LLMs, and so I can't say if its good or bad. I've played with them a little bit on my studio, but it became quite clear I need more ram then the base Studio came with

I’ve been building up my AI knowledge since getting my Studio.

My current goal is training my own local AI to help with my story-writing. The main hurdle is compiling a decent quality dataset (which is where I’m at now).

One interesting use I have for AI is that I use a python script to send a scene from my story to an LLM (via LM Studio’s API), get it to determine the main key moment of the scene, and create an image generation prompt. My script then sends that prompt to A1111 to generate the image which my script then inserts into the scene document so the scene has an illustration. This then continues for all of my scenes.

Over the last couple of days I’ve been running a “quality test” on my local LLMs. I have quite a few LLMs downloaded (>80) and I want to discover the best one for my story-writing tasks. So I set up a python script to send a full story (initially my shortest, at ~40,000 words) with a prompt to analyse it, respond with character journeys, key plot points, overview, etc. This repeats through all of my valid models (ones with a large enough context), allowing me to later read the responses and narrow down the models until I get to a shortlist, or a “winner”.

Other uses have been to get an LLM to go through my story scenes, create a “story so far” paragraph, then have that inserted into the next scene’s YAML frontmatter so that, when the next scene is processed in the same way, it can retrieve the “story so far” first for additional context.

Bringing together two of my hobbies like this has been good for me. I’m enjoying myself enormously.

 

[Tdude96]

I’ve been building up my AI knowledge since getting my Studio.

My current goal is training my own local AI to help with my story-writing. The main hurdle is compiling a decent quality dataset (which is where I’m at now).

One interesting use I have for AI is that I use a python script to send a scene from my story to an LLM (via LM Studio’s API), get it to determine the main key moment of the scene, and create an image generation prompt. My script then sends that prompt to A1111 to generate the image which my script then inserts into the scene document so the scene has an illustration. This then continues for all of my scenes.

Over the last couple of days I’ve been running a “quality test” on my local LLMs. I have quite a few LLMs downloaded (>80) and I want to discover the best one for my story-writing tasks. So I set up a python script to send a full story (initially my shortest, at ~40,000 words) with a prompt to analyse it, respond with character journeys, key plot points, overview, etc. This repeats through all of my valid models (ones with a large enough context), allowing me to later read the responses and narrow down the models until I get to a shortlist, or a “winner”.

Other uses have been to get an LLM to go through my story scenes, create a “story so far” paragraph, then have that inserted into the next scene’s YAML frontmatter so that, when the next scene is processed in the same way, it can retrieve the “story so far” first for additional context.

Bringing together two of my hobbies like this has been good for me. I’m enjoying myself enormously.

Sounds like you're making great use of it! What model(s) are you finding best for long context tasks?

 

[JSRinUK]

Sounds like you're making great use of it! What model(s) are you finding best for long context tasks?

My latest test has just finished - asking for an analysis and breakdown of my ~173,000 word story.

The only model to make it through my tests, and that has sufficient context length for this, is qwen3-vl-30b-a3b-instruct-mlx. This surprised me because it’s a vision language model, I believe, and these aren’t usually so great on text inference.

Almost all of the Qwen models scored high in my tests. Prior to this test, I’d been leaning into Gemma for my story-writing brainstorming (Gemma ones seem to be less prudish, and there are some uncensored ones out there that have been helpful, and they are pretty creative compared to other models), but Qwen has just been knocking it out of the park in terms of coherence and attention over long context.

I was ranking the models outputs on a score of 1-9 (with 9 being the best). I’d given a couple of the Qwen models a ‘9’ (“excellent, with some minor inaccuracies”), but then I read the output from qwen3-vl-30b-a3b-instruct-mlx and it just blew me away. It got the only “10” on my 9-point scale.

I now need to go and read the ~2,000 word output it’s given me for my ~173,000 word story. I’m hoping it didn’t lose anything with the extra long context, but even if it does - it still wins because it championed the ~37,000 word story and the ~62,000 word story.

 

[JSRinUK]

My latest test has just finished - asking for an analysis and breakdown of my ~173,000 word story.

The only model to make it through my tests, and that has sufficient context length for this, is qwen3-vl-30b-a3b-instruct-mlx. This surprised me because it’s a vision language model, I believe, and these aren’t usually so great on text inference.

Almost all of the Qwen models scored high in my tests. Prior to this test, I’d been leaning into Gemma for my story-writing brainstorming (Gemma ones seem to be less prudish, and there are some uncensored ones out there that have been helpful, and they are pretty creative compared to other models), but Qwen has just been knocking it out of the park in terms of coherence and attention over long context.

I was ranking the models outputs on a score of 1-9 (with 9 being the best). I’d given a couple of the Qwen models a ‘9’ (“excellent, with some minor inaccuracies”), but then I read the output from qwen3-vl-30b-a3b-instruct-mlx and it just blew me away. It got the only “10” on my 9-point scale.

I now need to go and read the ~2,000 word output it’s given me for my ~173,000 word story. I’m hoping it didn’t lose anything with the extra long context, but even if it does - it still wins because it championed the ~37,000 word story and the ~62,000 word story.

Following up - I’ve just read the 2,000 word output from qwen3-vl-30b-a3b-instruct-mlx after it analysed my longer 173,000 words story.

Conclusion: just “wow!”

It pretty much covered everything important in the story, discussing everything clearly and concisely. There is, perhaps, if I’m being hyper-critical, one tiny error/misunderstanding in the response, but that’s really a nothing. It doesn’t include everything, or every character, or every sub-plot, but that would have taken 10,000 words and I limited it to 3,000 tokens. It’s definitely enough, though.

Definitely a winner for my purposes.

This exercise has opened my eyes to the fact that not all models are made the same. That’s a good thing, because there are different models for different people.

But I know which one I’ll be using in the future.