DGX Spark 64GB: what published tests can tell us
The new 64GB DGX Spark asks buyers to make a fairly specific trade: enter NVIDIA’s compact AI platform with less memory than the original 128GB machine. The immediate question is how that changes the workloads it can serve.
Research checked 31 October 2026. Compiled from published tests. No hardware benchmarks were run for this article; measurements belong to the credited testers.
NVIDIA announced partner-built 64GB systems for October 23, starting at $4,999. The company says they retain GB10, DGX OS and the NVIDIA AI software stack. Its announcement reports up to a 1.7× performance improvement from pairing two 64GB systems in a Qwen3.8-27B test. Those are launch details and a vendor measurement, not an independent purchasing verdict. NVIDIA’s 64GB announcement.
The October 21 software release also adds support for both memory configurations, headless Wi-Fi setup and an option to reduce display-reserved memory on systems without a connected display. These are practical ownership improvements, but they do not establish model throughput. NVIDIA’s October release notes.
What existing tests can tell us
Ollama’s October 23, 2025 testing of the original Spark illustrates how much the model changes the experience. Using firmware 580.95.05 and Ollama 0.12.6, with ten repetitions, caching disabled and a 500-token output limit, the team reported:
| Model | Quantization | Generation, tokens/s |
|---|---|---|
| GPT-OSS 120B | MXFP4 | 41.14 |
| Llama 3.1 70B | Q4_K_M | 4.423 |
These are measurements of the older, larger-memory system. They show why parameter count alone is a poor speed predictor; they do not prove that either workload fits or performs the same on a 64GB Spark. Ollama’s test and reproducible script.
Physical implementation matters too. In his December 26, 2025 review of Dell’s 128GB GB10 machine, Jeff Geerling reported a different cooling arrangement and a larger power supply than NVIDIA’s original Spark. Dell supplied the systems; Geerling disclosed that it did not pay for or control the review. His observations concern that Dell model, so they should not be transferred to every new partner-built 64GB system. Jeff Geerling’s Dell review.
What independent tests still need to establish
The sources retrieved for this article did not include a detailed independent 64GB review documenting exact model files, context limits, sustained performance and power measurements. That limits how confidently anyone can recommend the new version from this research.
An independent review needs to establish usable memory after system overhead, performance as context grows, and what happens when requests overlap. A two-unit test also needs to include the cost and configuration of the connection between them.
Until then, the 64GB Spark is best discussed as a newly announced configuration with a defined software ecosystem. Old 128GB results provide background, not a substitute for testing the new machine.
Read the series
- Mac Studio and RTX cards: reading published LLM tests
- DGX Spark 64GB: what published tests can tell us (this article)
- AMD local LLM tests: the configuration behind the score
- Why local LLM benchmark charts disagree
- Local AI cost calculations: power, usage and payback
For memory planning, use the model-fit guide. Browse all Local AI notes.
Use this
Check if a model fits →
← Previous: Mac Studio and RTX cards: reading published LLM tests