NVIDIA DGX Spark 64 GB Reduces Memory to Stem RAM Price Increases
NVIDIA says the 64GB version of the DGX Spark retains all the other features of its older sibling, including the ConnectX7 LAN chip with dual QSFP ports good for 200 Gbps of networking. That's going to come in clutch for 64 GB system buyers, as bigger models will necessitate more systems on which to run them. It's not half a DGX Spark; it's a DGX Spark-like system with half the memory. These 64GB models will be available exclusively from NVIDIA's partners in the form of pared-back versions of existing units, like the Dell Pro Max with GB10 for $4,999. Other partners listed on NVIDIA's site include MSI, ASUS, Gigabyte, and Acer.
You might have noticed that $4,999 is more than $4,699, and thus be asking yourself what happens with the 128 GB version of the DGX Spark? That's still around, but the price has increased significantly to $6,995. NVIDIA's website might still be holding onto a cached price as occasionally we still see the old price (captured for posterity) but that puts its price in line with most partner systems including the HP ZGX Nano G1n, the review for which we just published.
On the AMD side, there are still several systems available at more affordable prices. In particular, GMKtec's AI Mini PC based on the Ryzen AI Max+ 395-based AI is available with 128 GB of RAM still for $3,650, while the 64 GB version is just $2,349. While our testing with the AMD Ryzen AI Halo (which was just freshened up for the HP system's review) showed that it wasn't as spry as GB10-based systems, the price difference may justify its performance, and a speedy cluster of these would set you back much less money. We do expect that AMD systems will see a similar jump in price, so if your workload is okay on red team hardware, now is as good of time as any to make a jump.
It's always painful to see an existing product get a big spec cut, but it's helpful to be mindful of what will run on such a system. First of all, the most popular models including Qwen 3.8 27B and Nemotron 3.5 Lightning 30B-A3B will still work on 64GB of RAM. It's important to recognize that, frankly, mid-sized dense models don't run that quickly on 256-bit LPDDR5x, although hybrid mixture-of-experts models like Nemotron 3.5 do zip right along. NVIDIA's blog "helpfully" points out that this is still sufficient memory for most agents to do things like summarization and autonomous computer usage. The fact that most of those models fit onto (and perform better on) 16GB GPUs is excluded from the announcement, however.
The big loss here for AI aficionados comes in the form of models that weigh in around 80B parameters or more, including Qwen 3.8 Flash Next, which we can run on the ZGX Nano. However, that model demonstrated that Alibaba and other open-weight makers are working to offload the lightweight parts of the model to storage. For example, Qwen 3.8 Flash Next has a 51B "parameter" phrase dictionary that gets read right from storage along with its 81B-parameter LLM. As that kind of split works its way into smaller models, a 64GB system will benefit.
The new 64GB GB10 systems will be available on October 23, according to NVIDIA's GB10 64 GB blog post announcing the systems. There's no word on whether the storage for these units will be reduced in kind, but we did find that 1 TB is manageable as long as you're not downloading a ton of models without curating what's on the device. Head on over and take a gander at what will be available.
