For most people running local AI today, yes, 128 GB is still enough, but the ceiling is getting lower relative to the models being released. A 128 GB machine comfortably holds the 100–120B mixture-of-experts class at 4-bit, such as gpt-oss-120b (about 65 GB) or Qwen3.5-122B-A10B (about 70 GB). It cannot hold the new frontier of open models, which now runs from about 750 billion to 2.8 trillion parameters. That is where the two platforms split. Every NVIDIA RTX Spark system, including the ASUS ProArt GR1X, tops out at 128 GB, and by default only part of it goes to the GPU. The Mac Studio M5 Ultra starts at 96 GB and goes to 256 GB or 512 GB. So the real question is whether your models will stay in the 128 GB class for the life of the machine.
What fits in 128 GB, 256 GB and 512 GB?
| Model (format) | Approx. size | RTX Spark 128 GB | Mac Studio M5 Ultra 256 GB | Mac Studio M5 Ultra 512 GB |
|---|---|---|---|---|
| gpt-oss-120b (native MXFP4) | ~65 GB | Yes, if the GPU share is raised above the default | Yes | Yes |
| Qwen3.5-122B-A10B (4-bit) | ~70 GB | Yes, with a raised GPU share | Yes | Yes |
| Llama 3.3 70B (FP8) | ~71 GB + context | Tight | Yes | Yes |
| Qwen3-235B (4-bit) | ~118 GB + overhead (our estimate) | No | Yes | Yes |
| Mistral Large 4 (very low-bit, ~1.5 bits) | ~233 GB | No | Yes (measured on a 256 GB unit) | Yes |
| DeepSeek R1 671B (4-bit) | ~404 GB | No | No | Yes, with GPU memory raised manually |
| Muse Spark Open 816B (NVFP4) | ~493 GB | No | No | No: Apple ran it on two or four linked Macs |
Source: model sizes from Hugging Face model files and Apple ML Research (15 Oct 2026); Mistral Large 4 on a 256 GB M5 Ultra from a community dataset reported by kingy.ai (Oct 2026); DeepSeek R1 on a 512 GB M3 Ultra from Dave2D via MacRumors (Mar 2025). The Qwen3-235B size is our calculation (235B parameters × ~0.5 bytes) and excludes context memory. Weights only unless stated; long conversations add more.
How much of that memory can the GPU actually use?
Neither platform hands all of its memory to AI by default, and this matters more than the headline number.
- RTX Spark (Windows on Arm): on 128 GB RTX Spark laptops, the default split gives the GPU about 64 GB, adjustable to about 96 GB in ASUS's MyASUS software (PCWorld, Oct 2026). On 64 GB units, Windows exposed at most about 55 GB to the GPU in XDA's testing. ASUS hasn't documented the setting for the GR1X desktop yet, so treat the laptop figures as a guide.
- Mac Studio (macOS): Metal's default GPU working set is about 75% of unified memory, roughly 72 GB on a 96 GB Mac, 192 GB on 256 GB and 384 GB on 512 GB. It can be raised with a system setting that Apple doesn't support, which is how 400 GB-class models were run on 512 GB M3 Ultra machines.
So a 128 GB RTX Spark realistically gives you somewhere between 64 GB and 96 GB for a model and its context, and a 256 GB Mac Studio gives you about 192 GB without tweaks. Our guide to why 128 GB of unified memory matters covers the basics of shared memory if you're new to it.
Why do open models keep outgrowing 128 GB?
Two trends pull in opposite directions. At the top end, open-weight releases have grown dramatically. Examples cited in our research include Kimi K3 at about 2.8 trillion parameters (roughly 1.4 TB) and GLM-5.2 at about 750 billion. One July 2026 analysis put the practical ceiling for a single local machine at the "100–120B MoE class at 4-bit; anything larger is a cluster or an API". Apple's own cluster demonstration ran an 816B model across two and four 512 GB Mac Studios precisely because it doesn't fit in one.
In the middle, labs are shipping smarter small models. NVIDIA's Nemotron 3.5 Bolt packs a 74B-parameter mixture-of-experts model into 24 GB using 2-bit quantisation, and mixture-of-experts designs only read a fraction of their parameters for each token. That keeps 128 GB useful for a lot of practical work: coding assistants, document question-answering and agents built on 30–120B models.
Context length adds its own pressure. A dense 70B model like Llama 3.3 needs about 0.33 MB of cache per token at 16-bit, which is roughly 43 GB at a 128K-token context on top of the weights (our calculation). Long documents and multiple users eat into a 128 GB machine faster than the model size suggests.
128 GB is enough for the models most people run today. It is not enough for the models that make headlines, and the gap is widening at the top.
Does more memory make the Mac faster?
Not by itself, but the M5 Ultra also has much more memory bandwidth, and that is what sets text-generation speed. Apple quotes 1.2 TB/s for the M5 Ultra. RTX Spark systems are around 300 GB/s (ASUS's figure; NVIDIA hasn't published one). On gpt-oss-120b, a same-chip RTX Spark laptop managed 63.3 tokens per second in Techtesters' October test, while a single community run on a 256 GB M5 Ultra reported about 110 tokens per second using a different engine. Those are not like-for-like numbers, but the direction matches the bandwidth gap.
Where RTX Spark keeps its edge is software and price. It runs CUDA, NVIDIA's software stack, although on Windows on Arm that is still a developer preview, and it costs less than a 256 GB Mac. For more on the Windows side, see our Windows on Arm app compatibility guide for RTX Spark.
What does each memory tier cost?
- RTX Spark at 128 GB: the ASUS ProArt P16 128 GB laptop is S$10,773.20 at SourceIT (listed 8 Nov 2026). The ProArt GR1X desktop isn't priced yet; a SourceIT estimate based on comparable RTX Spark devices, not an official ASUS price, puts the 128 GB/2TB model at about S$9,000–10,800.
- Mac Studio M5 Max 128 GB: about S$7,849 with a 1 TB SSD, calculated from Apple Singapore's upgrade pricing (confirm on apple.com/sg). Its 614 GB/s bandwidth sits between RTX Spark and the Ultra.
- Mac Studio M5 Ultra: Apple Singapore's list price starts at S$7,999 for 96 GB. A 256 GB/2TB configuration works out to about S$16,699, and 512 GB/4TB to about S$35,449, both calculated from Apple Singapore's upgrade pricing (confirm on apple.com/sg). The 256 GB and 512 GB options need the top 36-core CPU and 80-core GPU chip; there is no 128 GB Ultra.
Memory is fixed at purchase on both platforms, so you are effectively buying your model ceiling up front. We look at how that plays into lifespan in Mac Studio vs RTX Spark: which will be obsolete first.
So who needs more than 128 GB?
You probably don't if you run one model at a time in the 8B–120B range, mostly for a single user, with contexts up to about 32K tokens. RTX Spark, a GB10 box or an M5 Max Mac Studio all handle that.
You probably do if you want to run 235B-class or larger models locally, keep several models loaded at once, serve long contexts to more than a handful of people, or simply want headroom for the next two to three years of model growth. That is the case for a 256 GB or 512 GB M5 Ultra, or for NVIDIA systems with more GPU memory, such as an RTX PRO 6000 workstation. Our unified memory vs GPU power piece explains why capacity and speed are separate decisions, and whether unified memory replaces expensive GPUs looks at the trade-off from the GPU side. For a deeper look at how vendors quote model-size claims, see what NVIDIA's 120B and 200B claims really mean.
Buying in Singapore
SourceIT lists the Mac Studio M5 Ultra 256 GB/2TB and 512 GB/4TB configurations, plus the M5 Max 128 GB/1TB, as price on request; ask us for a quotation. On the RTX Spark side, the ASUS ProArt P16 128 GB is S$10,773.20 (listed 8 Nov 2026), and the ProArt GR1X desktop is available on request through our ProArt GR1X collection. MSI's RTX Spark laptops are in our MSI RTX Spark collection. Prices change, so ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can an RTX Spark PC use all 128 GB for AI models? Not by default. On 128 GB RTX Spark laptops the GPU gets about 64 GB out of the box, adjustable to about 96 GB in the maker's software. ASUS hasn't yet documented the setting for the GR1X desktop.
Is there a 128 GB Mac Studio M5 Ultra? No. The M5 Ultra comes with 96 GB, 256 GB or 512 GB. If you want 128 GB on a Mac Studio, that is the M5 Max, which has lower memory bandwidth than the Ultra.
Can a 512 GB Mac Studio run the biggest open models? It can run the DeepSeek 671B class at 4-bit with GPU memory raised manually. Larger models, such as the 816B model Apple demonstrated, needed two or more linked Mac Studios.
Does a longer context need more memory? Yes. For a dense 70B model, the context cache is about 43 GB at 128K tokens at 16-bit precision, on top of the model itself. Quantising the cache or shortening the context reduces it.
Is 128 GB a safe buy for the next three years? For single-user work on 30–120B models, it is likely to stay useful. If you expect to run 235B-class models, serve many users or keep several models loaded, buy more memory now, because neither platform lets you upgrade it later.
