# NVIDIA DGX Spark for local LLMs

> A quiet CUDA desktop with 128GB, but 273 GB/s limits speed and the price has climbed to about $6,950.

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).

## Key specs

- Memory: 128GB LPDDR5X (unified; ~119GB usable by the GPU by default)
- Memory bandwidth: 273 GB/s
- Architecture: Grace Blackwell (GB10) (2025)
- Power: 240W
- Typical price: ~$6,950 (checked 2026-10)
- Software: CUDA

## Best models for this GPU

- **Best overall (also coding, reasoning, vision):** Qwen3.6 35B-A3B at Q8_0, ~41.0GB with 8K context, ~32 tokens/sec
- **Runner-up (also largest that fits):** Qwen3.5 122B-A10B at Q6_K, ~104.5GB with 8K context, ~14 tokens/sec
- **Fastest good model:** gpt-oss-20b at MXFP4, ~15.0GB with 8K context, ~43 tokens/sec

## Compatibility (best quantization that fits, 8K context)

| Model | Params | Quant | Memory | Speed | Verdict |
| --- | --- | --- | --- | --- | --- |
| Gemma 4 E4B | 8B | Q8_0 | 10.5GB | ~21 tok/s | Runs well |
| Qwen3.5 9B | 9B | Q8_0 | 12.0GB | ~18 tok/s | Runs well |
| Phi-4 | 14.7B | Q8_0 | 18.3GB | ~13 tok/s | Runs well |
| gpt-oss-20b | 20.9B (3.6B active) | MXFP4 | 15.0GB | ~43 tok/s | Runs well |
| Gemma 4 26B A4B | 26B (3.8B active) | Q8_0 | 29.5GB | ~27 tok/s | Runs well |
| Qwen3.8 27B | 27B | Q8_0 | 31.6GB | ~6.9 tok/s | Runs well |
| Gemma 4 31B | 31B | Q8_0 | 35.9GB | ~6.2 tok/s | Runs well |
| Qwen3.6 35B-A3B | 35B (3B active) | Q8_0 | 41.0GB | ~32 tok/s | Runs well |
| GLM-4.5-Air | 106B (12B active) | Q6_K | 92.2GB | ~12 tok/s | Runs well |
| Llama 4 Scout | 109B (17B active) | Q6_K | 93.8GB | ~8.6 tok/s | Runs well |
| gpt-oss-120b | 117B (5.1B active) | MXFP4 | 68.3GB | ~37 tok/s | Runs well |
| Qwen3.5 122B-A10B | 122B (10B active) | Q6_K | 104.5GB | ~14 tok/s | Runs well |
| DeepSeek V4 Flash | 284B (13B active) | - | - | - | Won't fit |
| GLM-5.3-Flash | 320B (18B active) | - | - | - | Won't fit |
| GLM-5.3 | 744B (40B active) | - | - | - | Won't fit |

## Pros
- 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

## Cons
- 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

Full page: https://llamabuilds.ai/gpus/dgx-spark · Best models: https://llamabuilds.ai/gpus/dgx-spark/best-models

Sources: [NVIDIA DGX Spark product page](https://www.nvidia.com/en-us/products/workstations/dgx-spark/), [DGX Spark User Guide, hardware overview](https://docs.nvidia.com/dgx/dgx-spark/hardware.html), [IntuitionLabs DGX Spark review and pricing](https://intuitionlabs.ai/articles/nvidia-dgx-spark-review)

---
Source: https://llamabuilds.ai/gpus/dgx-spark · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.