NVIDIA · Grace Blackwell (GB10) · 2025
NVIDIA DGX Spark for local AI
A quiet CUDA desktop with 128GB, but 273 GB/s limits speed and the price has climbed to about $6,950.
Updated · Speeds marked ~ are estimates; seemethodology
- Unified memory
- 128GB
- Bandwidth
- 273 GB/s
- Power
- 240W
- Runs well
- 12 models
DGX Spark
Typically about $6,950
Can the DGX Spark run LLMs?
The DGX Spark has 128GB of unified memory (about 119GB usable by the GPU by default) and 273 GB/s of memory bandwidth, which decides how big a model fits and how fast it generates text. It comfortably runs models up to about 122B parameters: the largest in our catalog is Qwen3.5 122B-A10B at Q6_K (104GB with 8K context, ~14 tokens/sec). Our top all-round pick for it is Qwen3.6 35B-A3B (Q8_0, ~32 tokens/sec).
Recommendations
Best models for the DGX Spark
Best overall · also coding, reasoning, vision
Qwen3.6 35B-A3BQ8_0 · 41GB · ~32 tokens/sec
Runs wellRunner-up · also largest that fits
Qwen3.5 122B-A10BQ6_K · 104GB · ~14 tokens/sec
Runs wellFastest good model
gpt-oss-20bMXFP4 · 15GB · ~43 tokens/sec
Runs well
Compatibility
What runs on the DGX Spark
Every current model in our catalog, at the best quantization that fits in the ~119GB the GPU can use with 8K tokens of context. "Partial offload" means part of the model runs from system RAM (we assume 64GB), which is much slower.
| Model | Parameters | Best quant | Memory | Speed | Verdict |
|---|---|---|---|---|---|
| Gemma 4 E4B | 8B | Q8_0 | 10 GB | ~21 tok/s | Runs well |
| Qwen3.5 9B | 9B | Q8_0 | 12 GB | ~18 tok/s | Runs well |
| Phi-4 | 14.7B | Q8_0 | 18 GB | ~13 tok/s | Runs well |
| gpt-oss-20b | 20.9B (3.6B active) | MXFP4 | 15 GB | 86 tok/s | Runs well |
| Gemma 4 26B A4B | 26B (3.8B active) | Q8_0 | 30 GB | ~27 tok/s | Runs well |
| Qwen3.8 27B | 27B | Q8_0 | 32 GB | ~6.9 tok/s | Runs well |
| Gemma 4 31B | 31B | Q8_0 | 36 GB | ~6.2 tok/s | Runs well |
| Qwen3.6 35B-A3B | 35B (3B active) | Q8_0 | 41 GB | ~32 tok/s | Runs well |
| GLM-4.5-Air | 106B (12B active) | Q6_K | 92 GB | ~12 tok/s | Runs well |
| Llama 4 Scout | 109B (17B active) | Q6_K | 94 GB | ~8.6 tok/s | Runs well |
| gpt-oss-120b | 117B (5.1B active) | MXFP4 | 68 GB | 59 tok/s | Runs well |
| Qwen3.5 122B-A10B | 122B (10B active) | Q6_K | 104 GB | ~14 tok/s | Runs well |
| DeepSeek V4 Flash | 284B (13B active) | — | — | — | Won't fitWhat it needs |
| GLM-5.3-Flash | 320B (18B active) | — | — | — | Won't fitWhat it needs |
| GLM-5.3 | 744B (40B active) | — | — | — | Won't fitWhat it needs |
Performance
DGX Spark vs. alternatives on Qwen3.8 27B
Generation speed in tokens per second for a single user. Token generation is limited mostly by memory bandwidth, which is why bandwidth predicts these results so well.
RTX PRO 6000
Q8_0 · 96GB
~46tok/s
M5 Max 128GB
Q8_0 · 128GB
~13tok/s
DGX Spark
Q8_0 · 128GB
~6.9tok/s
Ryzen AI Max+ 395
Q8_0 · 128GB
~6.4tok/s
Measured
DGX Spark benchmark results
| Model | Quant | Runtime | Generation | Prompt | Source |
|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | llama.cpp | 86 tok/s | 3800 tok/s | llama.cpp discussion #16578: Performance of llama.cpp on NVIDIA DGX Spark · November 2025 |
| gpt-oss-20b | MXFP4 | llama.cpp (CUDA) | 83 tok/s | 4505 tok/s | llama.cpp repository: benches/dgx-spark/dgx-spark.md · February 2026 |
| Qwen3 Coder 30B-A3B Instruct | Q8_0 | llama.cpp (CUDA) | 61 tok/s | 2985 tok/s | llama.cpp repository: benches/dgx-spark/dgx-spark.md · February 2026 |
| gpt-oss-120b | MXFP4 | llama.cpp (CUDA) | 59 tok/s | 2445 tok/s | llama.cpp repository: benches/dgx-spark/dgx-spark.md · February 2026 |
| Llama 2 7B Chat | Q4_0 | llama.cpp (Vulkan) | 59 tok/s | 2790 tok/s | llama.cpp discussion #10879: Performance of llama.cpp with Vulkan · October 2026 |
| Llama 2 7B Chat | Q4_0 | llama.cpp (CUDA) | 57 tok/s | 3060 tok/s | llama.cpp discussion #15013: Performance of llama.cpp on NVIDIA CUDA · October 2026 |
Our take
Should you buy the DGX Spark for AI?
DGX Spark is NVIDIA's compact desktop built on the GB10 Grace Blackwell chip, with 128GB of LPDDR5X shared between a 20-core Arm CPU and a Blackwell GPU. Its 273 GB/s bandwidth is close to AMD's Strix Halo, so generation speed on large dense models is modest, but MoE models such as gpt-oss-120b run well and prompt processing is much faster than on Strix Halo. It ships with DGX OS and NVIDIA's AI software stack, and two units can be linked over the built-in ConnectX-7 networking. The Founders Edition launched at $3,999 in October 2025 and listed at $6,950 in October 2026.
What we like
- 128GB unified memory with full CUDA support
- Strong prompt processing for its size
- Two units can be clustered over ConnectX-7
- Small, quiet and low power
Watch out for
- 273 GB/s bandwidth limits token generation on dense models
- Arm Linux only, so some x86 tools need workarounds
- Price rose to about $6,950
Best for: Developers targeting NVIDIA's datacenter software stack · Running 70B to 120B MoE models quietly · Prototyping and fine-tuning with lots of memory
Specifications
DGX Spark specs that matter for AI
- Memory
- 128GB LPDDR5X (unified)
- GPU-usable memory
- ~119GB by default
- Memory bandwidth
- 273 GB/s
- Memory bus
- 256-bit
- Architecture
- Grace Blackwell (GB10)
- Power
- 240W (system)
- Interface
- Standalone system (ConnectX-7 for pairing two units)
- Released
- 2025
- Launch price
- $3,999
- Street price (new)
- ~$6,950 (2026-10)
- Software
- CUDA
Linux reports about 119GB, shared by the operating system and the GPU.
Sources: NVIDIA DGX Spark product page, DGX Spark User Guide, hardware overview, IntuitionLabs DGX Spark review and pricing
Cross-shop
Alternatives to the DGX Spark
Frequently asked questions
How much VRAM does the DGX Spark have?
The NVIDIA DGX Spark has 128GB of unified memory (about 119GB usable by the GPU by default) with 273 GB/s of memory bandwidth on a 256-bit bus.
What is the largest LLM the DGX Spark can run?
Fully in GPU memory, the largest model in our catalog that runs comfortably is Qwen3.5 122B-A10B (122B (10B active)) at Q6_K, using about 104GB with 8K context at ~14 tokens/sec.
Can the DGX Spark run Llama 3.3 70B?
Yes. At Q8_0 it needs about 81GB with 8K context and runs at ~2.7 tokens/sec.
What is the best LLM to run on the DGX Spark?
Our top pick is Qwen3.6 35B-A3B at Q8_0, which fits in 128GB with room for 8K of context and runs at ~32 tokens/sec. See our full list of the best models for the DGX Spark for coding, reasoning, and vision picks.
Is the DGX Spark good for local AI in 2026?
A quiet CUDA desktop with 128GB, but 273 GB/s limits speed and the price has climbed to about $6,950.