DGX Spark 64GB or 128GB: which one fits the models you run

NVIDIA's DGX Spark now comes in two sizes. The original has 128GB of unified memory. A 64GB version went on sale Oct 23, 2026 from Acer, ASUS, Dell, Gigabyte, HP and MSI, starting at $4,999, according to NVIDIA's announcement. Everything else about the two machines is the same, so the choice comes down to one question: how big are the models you want to run?

What's the same and what's different

Here's how the two compare on NVIDIA's spec page. Prices were checked Oct 28, 2026.

Both run DGX OS and the same NVIDIA software stack, and NVIDIA says runtimes such as llama.cpp, Ollama, vLLM and LM Studio work on both. Storage depends on the maker, up to 4TB. Because the 64GB model is sold only by partners, hardware support comes from whoever sold it to you.

How much memory your model needs

A quick rule from our VRAM guide: a model's weights take about its parameter count in billions, times the bits per weight, divided by 8, in gigabytes. A 70B model at 4-bit is about 35GB. Then add room for context (the KV cache), the runtime and the operating system, which share the same memory.

These are our estimates for the weights alone. For mixture-of-experts models, count total parameters, not active ones.

NVIDIA's ceilings of 100B and 200B parameters match roughly 4-bit weights filling most of the memory. In practice you'll want headroom. ASUS sizes its 64GB Ascent GX10 for models around 30 to 35B parameters, the range where you can keep long context and still have memory free. NVIDIA's own examples for the 64GB model include Nemotron 3.5 Bolt 74B, Qwen3.8 27B and Gemma 4 26B-A4B.

More memory doesn't mean more speed

Both versions read memory at 273 GB/s, and when a model runs locally, memory bandwidth mostly sets how fast it generates text. So a 27B model at 4-bit runs at about the same speed on the 64GB and the 128GB machine. What the 128GB buys you is room: bigger models, 8-bit instead of 4-bit, longer context, or two models loaded at once, such as a coding model next to a small one for quick edits.

Two 64GB units or one 128GB

Each Spark has two QSFP ports. NVIDIA says two 64GB units on one direct cable pool 128GB and handle models up to 200B parameters, with twice the memory bandwidth of one unit. In NVIDIA's test with Qwen 3.8 27B, the pair delivered up to 1.7x the performance. Its Sync Cluster Assistant sets the pair up.

The cost is the catch. At the starting price, two 64GB units come to $9,998, against $6,950 for NVIDIA's 128GB unit. A pair makes sense if you want the extra speed, or if you want to buy one machine now and add a second later. If you just need 128GB, one unit costs less. In its Oct 2 announcement, NVIDIA listed its playbook for connecting multiple Sparks as "coming soon" to 64GB devices, so check that it's out before you plan around a pair.

You can go bigger. NVIDIA's spec page says up to four systems: two or three on direct cables, four with a 200GbE switch. Two 128GB units reach 256GB, for models up to 400B parameters. NVIDIA's Oct 21 GTC Berlin post says clusters of up to eight units, or 1TB of memory, work in manual configuration. Treat anything beyond four as an advanced configuration.

When something else fits better

• You want a laptop. RTX Spark laptops have up to 128GB of unified memory, run Windows, and start at $2,599, though 128GB configurations cost more. NVIDIA says they handle models up to 120B parameters.
• Your models fit in 32GB. A desktop GPU such as the RTX 5090 has 32GB with far higher bandwidth, so it generates much faster on models that fit. Street prices have been well above its $1,999 launch price. Our unified memory comparison compares the options.
• You need more than 128GB in one box. A Mac Studio with M5 Ultra goes up to 512GB of unified memory. It doesn't have CUDA, though.

How to choose

1. Write down the biggest model you plan to run and the precision you'll run it at.
2. Work out its weights with the rule above, then add about 20% or more for context and the system.
3. If the total stays under about 45GB, the 64GB model is enough.
4. If it's between about 45GB and 110GB, or you want two models loaded at once, get the 128GB.
5. If it's beyond that, look at two 128GB units or a machine with more memory.