NVIDIA GB10 Grace Blackwell — Local LLM Performance & Compatibility
作者： Jakub Rusinowski · 最后更新： 2026年9月19日
DGX Spark及基于它的OEM机型内部的芯片——ASUS Ascent GX10、Dell Pro Max、HP ZGX Nano、Lenovo ThinkStation PGX。20核Grace ARM CPU与Blackwell GPU共享128 GB LPDDR5X，带宽273 GB/s，运行完整的CUDA栈。作为芯片单独记录，使这些机器可以共用一条记录，而不必各自重述。
Technical Specifications
| VRAM | 128 GB |
| Memory Bandwidth | 273 GB/s |
| TDP | 140 W |
| Architecture | GB10 Grace Blackwell Superchip |
| Release Year | 2025 |
| MSRP at Launch | $0 |
| Inference Speed (Llama 3.1 8B Q4_K_M) | 26–53 tok/s (estimated) |
| Inference Speed (Llama 3.3 70B Q4_K_M) | 3.4–7.1 tok/s (estimated) |
作为亚马逊联盟成员，我们从符合条件的购买中获得收入。云 GPU 链接为推荐链接——我们可能获得佣金，您无需额外付费。
LLMs Compatible with 128 GB VRAM
All models below run comfortably in 128 GB VRAM with Q4_K_M quantization.
| Qwen3.8 | Qwen3.8-Flash-Next · 109 GB VRAM · Q4_K_M · qwen3-8 |
| Qwen 3.5 | Qwen 3.5 122B-A10B · 74 GB VRAM · Q4_K_M · ollama run qwen3.5:122b |
| Mistral Small 4 | Mistral Small 4 119B-A6.5B · 73 GB VRAM · Q4_K_M · ollama run mistral-small |
| Llama 4 | Llama 4 Scout 17B · 67 GB VRAM · Q4_K_M · ollama run llama4:scout |
| Command R Family | Command R+ (104B) · 64 GB VRAM · Q4_K_M · ollama run command-r-plus |
| Llama 3.2 Family | Llama 3.2 90B Vision Instruct · 54 GB VRAM · Q4_K_M · llama-3-2 |
| Llama 3.2 Vision | Llama 3.2 Vision 90B · 54 GB VRAM · Q4_K_M · ollama run llama3.2-vision:90b |
| Qwen3-Coder | Qwen3-Coder 80B-A3B (MoE) · 49 GB VRAM · Q4_K_M · ollama run qwen3-coder:80b-a3b-q4 |
62 more families also fit 128 GB — browse the full model library.
Best Use Cases
- unified memory AI
- 70B models
- local development
FAQ
Can the NVIDIA GB10 Grace Blackwell run local LLMs?
Yes — the NVIDIA GB10 Grace Blackwell has 128 GB VRAM and runs DGX Spark及基于它的OEM机型内部的芯片——ASUS Ascent GX10、Dell Pro Max、HP ZGX Nano、Lenovo ThinkStation PGX。20核Grace ARM CPU与Blackwell G
How fast is the NVIDIA GB10 Grace Blackwell for AI inference?
The NVIDIA GB10 Grace Blackwell is estimated to run Llama 3.1 8B at 26–53 tok/s with Q4_K_M quantization. For Llama 3.3 70B the estimate is 3.4–7.1 tok/s. These are modelled estimates, not measurements — see /en/methodology.
What LLMs can I run on 128 GB VRAM?
With 128 GB you can run: Qwen3.8, Qwen 3.5, Mistral Small 4, Llama 4, Command R Family. Use Ollama for the easiest setup: ollama run llama3.1:8b.
Compare Similar GPUs
- Apple M4 Max (128 GB, 110 t/s)
- Apple M3 Max (128 GB, 95 t/s)
- Apple M1 Ultra (128 GB, 155 t/s)
- Apple M5 Max (128 GB, 125 t/s)
VRAM Tier
Buying Guide
← All GPU Reviews | All Hardware | Check Your Hardware | Full Benchmarks | Can I Run It?