# Shared memory lets local models exceed 128GB workloads

Markdown copy of https://realalpha.app/s/shared-memory-lets-local-models-exceed-128gb-workloads/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Ian Weinberg](https://realalpha.app/people/ian-weinberg/) · Ian Weinberg is an AMD representative.
- Host: [Sophia Dew](https://realalpha.app/people/sophia-dew/) · Host, MTS
- Show: MTS
- Episode: [Local AI Will Take Over Within One Year | Ian Weinberg](https://realalpha.app/v/PJlTf2ACCuY/), Oct 31, 2026
- Moment: 9:58 to 14:40 of the episode (5 min), playing on https://realalpha.app/s/shared-memory-lets-local-models-exceed-128gb-workloads/; the episode from there: https://realalpha.app/v/PJlTf2ACCuY/?t=598

## Quote

> “So right now, 128 plus 24, we are **outside of the envelope** of a traditional 128-bit system.”
> — [Ian Weinberg](https://realalpha.app/people/ian-weinberg/) at [13:44](https://realalpha.app/v/PJlTf2ACCuY/?t=824) · MTS

## Takeaways

1. A 176B hybrid model uses CPU memory for its n-grams.
2. The decode stage pushes dedicated GPU memory to 128GB.
3. Total memory use reaches 152GB, beyond a 128GB system.

## About this segment

Ian Weinberg runs a large hybrid model, then adds a second model that pushes memory use beyond 128GB. The demo shows why shared CPU and GPU memory matters for complex local workloads.

From “Local AI Will Take Over Within One Year | Ian Weinberg”, Oct 31, 2026. [Watch the full episode](https://realalpha.app/v/PJlTf2ACCuY/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[10:02](https://realalpha.app/v/PJlTf2ACCuY/?t=602) the thing that i'm going to show you today is um is think of creativity think about when you wake up in the morning and you're like am i a creative person and i will say i am not i'm artistic creative human being but i'm a very musically oriented person that's very cool very musically oriented so what i've been exploring is how do i start to create some music via prompting so there's a couple of things that are here and i recently went oh this is really cool okay so i know i get what you're saying when you're saying you wanted to design it something for something you specifically wanted to do okay something i wanted to do and by the way the next demo after this when we continue to talk it's going to be a self-running agentic demo that will happen right afterwards but this one is taking advantage of the quen 3.8 flash next so with that particular model

[10:52](https://realalpha.app/v/PJlTf2ACCuY/?t=652) it's um it's a hybrid model so it's got n grams and it's got traditional parameters and it's about 120 billion parameters plus another 56 or so um of of those n grams the n grams run in the cpe so that's something to understand but this is the kind of thing where you know i want to make me i want to make a song and it made a song for me and i don't love it so let's turn around and say hmm i like this but make it more like dance hall robin with a slower beat so effectively what i want it to do is think through how what kind of song is that i'm looking to build change the lyrics build it up however it's going to do and then i'm going to feed that into comfy ui and i'm very large model is running at you'll see it's running remind me the size of this model so this is a it's a hundred and seventy six I think total of the parameters they 120 million a billion of

[11:51](https://realalpha.app/v/PJlTf2ACCuY/?t=711) them are the per the traditional parameters and the rest of it is engrams which would just be look up on the CPUs yeah did you jump around so what you'll see here is this Loaded up, 98 gigs of memory. This is a cool task manager so you can see in real time. Isn't it fun? It's happening. But this is not the whole story because right now it's got 98 gigs of dedicated GPU plus 22 gigs of memory. So right now, add those two things together and where'd we get?

[12:15](https://realalpha.app/v/PJlTf2ACCuY/?t=735) We're like at 120, right? Yeah. Roughly in that range. Now, that would be great if that was all that we were doing. But if that was all that we were doing, then I could do that in a 128 gig system. Oh, fascinating. And are we going to demo and show what that might be? We're going to go past 128 gigs now. Oh. We're going to drop this into ComfyUI, and I'm going to have ComfyUI using another model make a song for me.

[12:41](https://realalpha.app/v/PJlTf2ACCuY/?t=761) And when it does that, what you will see happening is, so in this case, we use A-step 1.5. A-step 1.5. Yes. So from there, what I'll do is I'll have that kick off. And really, it's when you get to the decode that things get interesting. So you'll watch that it will start to build up. 105 and as we're moving through the prompting step and the end code yeah and give it a couple of seconds and what it's doing now is this coming up i don't know what the audio is going to work on this but if it does i could play it for you at the end yeah i think we have to share audio and but it's okay we can we'll hear it at the end we'll we'll get there towards the end so as it continues going through you're seeing more and more stuff being loaded into memory as it gets to the decode stage.

[13:26](https://realalpha.app/v/PJlTf2ACCuY/?t=806) Uh-oh, here we go. That's going to be real hard to do 128 gigs of dedicated GPU. Can you tell me what's happening right now in the task manager? Yeah, so right now in the task manager, you'll see dedicated GPU memory has gone to 128 gigs. But you have to add the memory that's being used by the system to that. So right now, 128 plus 24, we are outside of the envelope of a traditional 128-bit system.

[13:53](https://realalpha.app/v/PJlTf2ACCuY/?t=833) Yeah, I'm just trying to imagine running this on my MacBook. So I'll let the other guys sell their platforms, but what I will say right now is the only platform I could do this on is this one. Wow, yeah. Even the other competitors. So it gives you an idea of what this unlocks. It unlocks the ability to go to much more complex models, to do more complex reasoning, to be able to take on tasks that traditionally would not have been thought of as being local.

[14:17](https://realalpha.app/v/PJlTf2ACCuY/?t=857) Yeah, that's impressive. I think one of the things that I'm particularly excited about is the connection between the software and the hardware. I haven't particularly seen that before. I love playing around with these different hardwares, but you don't see it straight out of the box. You're often having to figure that out on your own, and then you don't know what's actually optimized for the hardware that you're trying to plug into. So this is really fascinating to see.