Compact AI development desktop built around the GB10 Grace Blackwell Superchip — a 20-core Grace ARM CPU plus a Blackwell GPU (6,144 CUDA cores) sharing 128 GB of unified LPDDR5x memory at 273 GB/s. Runs the full CUDA stack, unlike Apple Silicon, making it ideal for fine-tuning and inference of 70B-class models. Launched at $3,999 in October 2025; price raised to $4,699 in February 2026 due to memory supply constraints.
| VRAM | 128 GB |
| Memory Bandwidth | 273 GB/s |
| TDP | 150 W |
| Architecture | GB10 Grace Blackwell Superchip |
| Release Year | 2025 |
| MSRP at Launch | $4,699 |
| Inference Speed (Llama 3.1 8B Q4_K_M) | ~60 tokens/sec |
| Inference Speed (Llama 3.3 70B Q4_K_M) | ~20 tokens/sec |
All models below run comfortably in 128 GB VRAM with Q4_K_M quantization.
| Llama 3.3 | 24 GB VRAM · Q2_K_XS (Tight) · ollama run llama3.3 |
| Llama 3.1 Family | 6 GB VRAM · Q4_K_M · ollama run llama3.1 |
| Llama 4 | 24 GB VRAM · Q4_K_M · ollama run llama4:maverick |
| DeepSeek R1 | 20 GB VRAM · Q4_K_M · ollama run deepseek-r1:32b |
| Qwen 3 | 80 GB VRAM · Q4_K_M · ollama run qwen3:235b-a22b |
| Qwen3-Coder | 8 GB VRAM · Q4_K_M · ollama run qwen3-coder:80b-a3b-q4 |
| Gemma 3 | 16 GB VRAM · Q4_K_M · ollama run gemma3:27b |
| Mistral Small 3.1 | 14 GB VRAM · Q4_K_M · ollama run mistral-small3.1 |
Install Ollama then run the recommended model for this GPU:
ollama run llama3.3:70b
Yes — the NVIDIA DGX Spark has 128 GB VRAM and runs Compact AI development desktop built around the GB10 Grace Blackwell Superchip — a 20-core Grace ARM CPU plus a Blackwel
The NVIDIA DGX Spark runs Llama 3.1 8B at ~60 tokens/sec with Q4_K_M quantization. For the 70B model it achieves ~20 tokens/sec.
With 128 GB you can run: Llama 3.3, Llama 3.1 Family, Llama 4, DeepSeek R1, Qwen 3. Use Ollama for the easiest setup: ollama run llama3.3:70b.
← All GPU Reviews | Check Your Hardware | Full Benchmarks | Can I Run It?