NVIDIA DGX Spark for LLMs: what it runs
Updated 2026-11-03
The NVIDIA DGX Spark is a small desktop computer built around the GB10 Grace Blackwell chip with 128 GB of memory shared by the CPU and GPU. It went on sale in October 2025 at a launch price of $3,999. For companies looking at private AI it answers one question well: can we run a serious open-weight model on our own desk, without a server room?
Specifications that matter for LLMs
| DGX Spark | |
|---|---|
| Chip | NVIDIA GB10 Grace Blackwell (20-core Arm CPU + Blackwell GPU) |
| Memory | 128 GB LPDDR5x, unified (shared by CPU and GPU) |
| Memory bandwidth | about 273 GB/s |
| AI compute | up to 1 petaflop at FP4 (with sparsity) |
| Storage | up to 4 TB NVMe |
| Networking | ConnectX-7, to link two units |
| Software | DGX OS (Ubuntu-based) with the NVIDIA AI software stack |
NVIDIA positions one unit for models up to about 200 billion parameters at 4-bit, and two linked units for about 405 billion.
Which models fit in 128 GB
Memory is the first limit. Model weights take about 0.56 GB per billion parameters at 4-bit, 1 GB at 8-bit and 2 GB at 16-bit, plus the KV cache for the text being processed (use the LLM VRAM calculator for your case).
| Model class | 4-bit weights | Fits on one DGX Spark? |
|---|---|---|
| 8B | about 4.5 GB | Yes, even at 16-bit |
| 32B | about 18 GB | Yes, with room for long context |
| 70B | about 40 GB | Yes at 4-bit or 8-bit |
| 120B mixture-of-experts (gpt-oss-120b class) | about 66 GB | Yes |
| 405B | about 228 GB | No; needs two linked units at lower precision, or a server |
How fast it generates
The second limit is speed. When a model writes an answer, every new token reads all the active weights from memory once, so for a single request the upper bound on speed is roughly:
tokens per second ≈ memory bandwidth ÷ bytes of active weights
With about 273 GB/s, a dense 70B model at 4-bit (about 40 GB) tops out near 7 tokens per second for one request; real numbers are lower. A 32B model at 4-bit has a ceiling near 15 tokens per second. Mixture-of-experts models are faster than their size suggests, because only a few experts run per token: gpt-oss-120b activates about 5 billion parameters per token, so it generates far faster than a dense 70B model while needing more memory.
For comparison, datacenter GPUs have several times the bandwidth: an H100 has about 3.35 TB/s and an H200 about 4.8 TB/s. That is why a GPU server serves many users at once, while a DGX Spark suits one person or a small team.
When a DGX Spark makes sense for a company
- A pilot. Test open-weight models on your own documents, inside your network, before buying servers.
- One team or one workflow with modest traffic, such as document extraction overnight or an assistant for a handful of people.
- Development. Engineers build and fine-tune small models locally, then deploy the same software stack on servers.
- Data that cannot leave a room, where a desktop machine is easier to approve than cloud access.
When you need a GPU server instead
- Dozens of people using the assistant at the same time.
- Interactive chat on 70B-class dense models, where people expect answers in a few seconds.
- Long documents for many users at once, where the KV cache grows past what is left after the weights.
- Uptime requirements: a server with redundant power, monitoring and a second machine for failover.
The LLM cost calculator shows when owned hardware beats API pricing at your volume; our hardware guide compares the DGX Spark with workstation GPUs, Macs and servers.