Nvidia has introduced a lower-memory version of its DGX Spark AI mini PC, the DGX Spark 64 Go, which features half the memory of the previously launched 128 Go model. The 64 Go version is priced at 4,999 dollars and will be available from October 23 exclusively at Acer, Asus, Dell, Gigabyte, HP, and MSI. Nvidia will not offer a 64 Go Golden Founders Edition. The 128 Go model, launched at 3,999 dollars in October 2025, has seen its price increase significantly, reaching 6,950 dollars by October 2, a 74% increase in one year. In France, the Founders Edition had gone from 4,099 to 4,800 euros in February, and the Dell Pro Max 4 To version tested was already at 7,149 euros at Dell before this new increase. The European price for the 64 Go version is not communicated, but it is expected to be around 5,000 euros, with additional costs from manufacturers for on-site warranty. The DGX Spark 64 Go uses the same GB10 Grace Blackwell chip, 20-core Arm CPU, 6,144-core Blackwell GPU (equivalent to an RTX 5070), 273 Go/s memory bandwidth, and ConnectX-7 network card as the 128 Go model. The only differences are the reduced RAM (64 Go instead of 128) and the SSD (2 To targeted instead of 4). Nvidia states that the 64 Go version is the same circuit board with fewer memory chips soldered on it to stay under the 5,000 dollar threshold. However, the 4,999 dollars price remains 25% more expensive than the 128 Go model from twelve months ago. The DGX Spark 64 Go is not a new machine but a scaled-down version of the 128 Go model. Nvidia mentions that models up to 100 billion parameters can be run on a single unit, assuming aggressive quantization and reasonable context usage. However, in practice, a dense model of 27 to 32 billion parameters or a mixture-of-experts type Qwen3.8 will be comfortable. A 70B quality model would be much less. The Qwen3.5-122B-A10B model tested occupied 85 Go of memory, which does not fit into a Spark 64 Go. The 128 Go model allowed loading models that have little to do with those of a consumer graphics card, while the 64 Go brings the machine back into the realm of "average" models. The 273 Go/s bandwidth is the real bottleneck. Qwen3.5-122B had difficulty exceeding 20 tokens per second, and a small Qwen3.5-14B did not go faster. On these small models, a GeForce RTX 5060 Ti with 16 Go, sold around 850 euros, is almost twice as fast. The GB10 is much more comfortable in image generation: on Flux.1 Dev, it clearly exceeds the RTX 5060 Ti and approaches an RTX 3090, silently and without heating the room. Fine-tuning works (a LoRA Flux.1 Dev in three hours without worrying about fitting the model in memory), but it is the usage that will suffer the most from half the RAM less. Nvidia admits that for training, the 128 Go remains the right choice. It is not an inference server. Serving a large model to multiple users at the same time is not feasible, and a long conversation of 65,000 tokens made the response wait up to three minutes. DGX OS, the homemade Ubuntu, is guaranteed up to date only two years. On a 5,000 euro machine, it is short. Nvidia's central argument is the cluster. Two Spark 64 Go units connected by a QSFP cable on their ConnectX-7 port (200 Gb/s) add up to 128 Go of memory, double the computing power and double the bandwidth. On Qwen3.8 27B, the company measures up to 70% more performance compared to a single 128 Go. The figure is credible, but it still deserves to be looked into. This model fits in a single 64 Go box, the gain comes from the doubling of the bandwidth. Two Spark 128 Go units would do the same thing, with 256 Go. And the addition follows: 9,998 dollars for two units, 44% more than a single 128 Go, two power supplies, two 2 To SSDs to fill and 280 W on the desk. The Nvidia Sync application includes a cluster assistant that detects the units and configures the network without touching a file, and a "Model Launcher" arrives at the end of October to launch Qwen3.8 27B on one or two machines with a few clicks, with the OpenCode editor connected. Last subtlety: mixing a 64 Go and a 128 Go works, but the model will be split on 64 Go per machine. The remaining 64 Go of the large model will sleep. Facing AMD, Apple, and Nvidia itself, the DGX Spark 64 Go has competition. AMD first, with its Ryzen AI Max+ 395 (Strix Halo) found in a dozen mini PCs at GMKtec, Minisforum, Beelink, Bosgame, Framework, or HP. Same type of LPDDR5X memory, same 273 Go/s bandwidth, so about the same token generation speed on a given model. A GMKtec EVO-X2 in 64 Go is traded around 2,000 euros on the European store of the brand, a Framework Desktop 64 Go without SSD starts under 2,000 dollars, a Minisforum MS-S1 Max 64 Go with 2 To runs around 2,600 dollars. They also took 50 to 70% in a year because of memory, but they remain twice as cheap as the Spark. What we lose by going to AMD: CUDA, the pre-installed Nvidia software stack (it was the big point of our test, everything runs in a few minutes with Docker and the NIM containers), the 200 Gb/s port for the cluster, and part of the raw computing power, the GB10 Blackwell GPU still ahead of the Radeon 8060S on optimized loads. What we gain: Windows or Linux at choice, a machine that also serves other purposes than AI, and 2,500 euros saved. Apple is also in the battle. The Mac Studio M5 Max starts at 2,999 euros with 36 Go, and its memory bandwidth exceeds the 273 Go/s of the GB10 by a large margin, which is directly felt in the generation speed. The M5 Ultra, starting at 6,599 euros with 96 Go, goes up to 512 Go and 1.2 To/s. In other words, at the price of a DGX Spark 128 Go, Apple sells a machine with almost as much memory and four times more throughput. The downside: no CUDA, an inference ecosystem (MLX, llama.cpp) less equipped for production, and memory options that Apple has also raised this year. Even the Mac mini M5 Pro, at 3,329 euros in 64 Go, comes to walk on the grounds of the Spark 64 Go, with less GPU power but a real desktop computer around. The most annoying competitor could come from Nvidia itself. At Computex, end of May, the brand announced RTX Spark, its chip for Windows PCs. Under the marketing name, we find the same GB10 as in the DGX Spark, available in N1X (the full GB10) and N1 (a lighter variant), with configurations of 16 to 128 Go at Acer, Asus, Dell, HP, Lenovo, MSI, and Microsoft Surface. Asus has already shown its ProArt GR1X, a silent RTX Spark mini PC designed to run agents 24/7, exactly the argument of the DGX Spark 64 Go. The first machines are expected this autumn, but there are no official prices yet. Morgan Stanley estimates 2,899 dollars for an N1X and 1,799 dollars for an N1. If these estimates hold, a 64 Go RTX Spark PC under Windows would cost less than a DGX Spark 64 Go under Linux, with the same chip. The difference will be on the software stack (DGX OS and its ready-to-use containers against Windows and its Prism emulation layer) and on the probable absence of the ConnectX-7 port on the consumer PC side. Nvidia has indeed confirmed that there will be no Windows on the DGX Spark: the two ranges will remain separate.