Product1 publisher3 min readPublished
Nvidia adds a 64GB DGX Spark sized for 30B-class local models
Nvidia's DGX Spark 64GB ships October 23 from Acer, ASUS, Dell, Gigabyte, HP and MSI, with the same GB10 chip as the 128GB model and half its memory. Teams comparing it with cloud token bills now start by asking how much of their workload fits in 64GB.
The Product Desk · Product desk
What happened
- Nvidia aims the 64GB tier at 30-35B class models and autonomous agents.
- The GB10 pairs a Blackwell GPU rated at up to one petaflop of FP4 AI compute with a 20-core Grace Arm CPU, both drawing on one LPDDR5X memory pool.
- NVIDIA says its NVFP4 format lets the Spark run compressed versions of larger models without a loss in accuracy.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision A team that outgrows 64GB has to choose between one 128GB unit and adding a second 64GB unit, and price per gigabyte from the OEM quotes will decide it.
- cost The pitch swaps a per-token cloud bill for an upfront hardware purchase, so the box pays back only for teams whose model use is steady enough to keep it busy.
- exposure Teams counting on NVFP4 compression to fit bigger models into 64GB are relying on Nvidia's own accuracy claim until independent tests exist.
Picture whoever files the purchase request for one of these next week. They have six vendor names and a ship date [1]. The unit-price field stays empty. PCMag's launch article says the 64GB configuration comes "without requiring an enterprise-grade price tag" [4], but it does not state a price [13].
Most of the article is a list of jobs the box could do. It could run agents that manage email and smart home devices, handle vibe-coding without buying more tokens, fine-tune a model on your own source code, and run long agents "without a cloud meter ticking in the background" [9]. The product itself is a memory configuration. It is the same GB10 Superchip, with its CPU and GPU sharing 64GB of LPDDR5X where the original had 128GB [2][5][11].
Every job on that list draws from one pool. The article says the shared memory lets multiple models and agents work side by side and hold long-running context [8]. On a 64GB unit, model weights, the context for long tasks and any second agent all compete for the same 64GB. For a team, the useful question is the size of the largest model it actually runs, at the precision it runs it, plus the context it keeps. I'd size the purchase from a workload the team runs and measures today. The email-managing agent can stay out of the spreadsheet until somebody builds one.
Nvidia's stated target is 30-35B class models [3]. For scale, the article says many current AI PCs ship with 32GB of RAM and integrated GPUs that fall short on LLM inference [10]. The Spark 64GB has twice that memory [1]. Beyond that, NVIDIA says its NVFP4 format lets the Spark run compressed versions of larger models without a loss in accuracy [7]. That accuracy claim comes from Nvidia, passed on by PCMag. A team planning to squeeze a bigger model into 64GB through compression is taking it on trust for now.
If a team outgrows the box, the next step is a second box. ConnectX-7 networking is built in, and two 64GB units can be clustered to reach 128GB with added compute [6]. That is the memory the original Spark launched with at CES 2025 [11]. So the 64GB unit is a sensible first purchase for a team unsure of its ceiling, as long as two small units do not cost more than one large one. The OEM quotes will answer that.
The decision fits on a 2x2. One axis is whether the largest model you run, at working precision and with its context, fits in 64GB. The other is whether your cloud token spend is steady from month to month or comes in bursts.
A model that fits, paired with steady spend, is the case this unit is built for. Payback is the box price divided by the monthly cloud bill it replaces. A model that fits but bursty spend leaves the box idle between bursts, and the meter may well cost less. When the model does not fit and spend is steady, compare one 128GB unit with two clustered 64GB units on price per gigabyte. When it does not fit and spend is bursty, stay in the cloud.
What to watch
- Price lists from Acer, ASUS, Dell, Gigabyte, HP and MSI, and whether two 64GB units cost more or less than one 128GB unit.
- Independent benchmarks of NVFP4-compressed models on the 64GB unit, tested against Nvidia's claim of no accuracy loss.
- Whether a clustered pair of 64GB units matches a single 128GB unit on time to first token and tokens per second.