NVIDIA GB10 Grace Blackwell for local LLMs
Written by Jakub Rusinowski · Last updated
With 128 GB of LPDDR5X at 273 GB/s, the GB10 Grace Blackwell runs 144 catalogued models at Q4_K_M with 8K context. The largest that fits is Step 3.7 Flash 198B-A11B (~119.7 GB), and the top pick is Qwen3.8-Flash-Next at 35–72 tok/s.
The silicon inside DGX Spark and the OEM boxes built on it — ASUS Ascent GX10, Dell Pro Max, HP ZGX Nano, Lenovo ThinkStation PGX. A 20-core Grace ARM CPU and a Blackwell GPU sharing 128 GB of LPDDR5X at 273 GB/s, running the full CUDA stack. Recorded as a CHIP so those machines can share one record instead of restating it.
Models that run on the GB10 Grace Blackwell
Q4_K_M, 8K context, 128 GB usable. Ranked by quality and speed.
| Model | Memory · marker = 128 GB | VRAM | Speed |
|---|---|---|---|
| Qwen3.8-Flash-Next Qwen3.8 | 109.7 GB | 35–72 tok/s | |
| Qwen 3.5 122B-A10B Qwen 3.5 | 74.5 GB | 16–33 tok/s | |
| Nemotron 3 Super 120B-A12B Nemotron 3 Super | 73.3 GB | 14–30 tok/s | |
| Nemotron 3.5 Super VL 120B-A12B Nemotron 3.5 | 73.3 GB | 14–30 tok/s | |
| Mistral Small 4 119B-A6.5B Mistral Small 4 | 72.6 GB | 21–43 tok/s | |
| GPT-OSS 120B GPT-OSS | 71.6 GB | 38–80 tok/s | |
| Llama 4 Scout 17B Llama 4 | 68.2 GB | 12–25 tok/s | |
| Llama 4.5 Scout Llama 4.5 | 66.6 GB | 11–23 tok/s | |
| Qwen3-Coder-Next (80B-A3B MoE) Qwen3-Coder | 49.4 GB | 58–120 tok/s | |
| Kolibri 1 Kolibri | 48.5 GB | 52–109 tok/s | |
| Nemotron 3.5 Bolt 74B Nemotron 3.5 | 45.6 GB | 25–52 tok/s | |
| K2 Horizon MoVA 36B-A4B K2 Horizon | 25 GB | 33–69 tok/s |
Buy it or rent the same memory
Buy the card, or rent a GPU with the same memory by the hour to try models first.
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
Speed vs other GPUs
Llama 3.1 8B, Q4_K_M. Estimated ranges. How this is calculated
Specifications
Specs last updated 2026-09-19.
- Memory
- 128 GB LPDDR5X
- Memory bandwidth
- 273 GB/s
- Memory bus
- 256-bit
- Architecture
- GB10 Grace Blackwell Superchip
- Series
- Grace Blackwell
- Board power
- 140 W
- Release year
- 2025
- Compute backends
- CUDA
- Usable for models
- 128 GB
Similar GPUs
Frequently asked questions
Can the NVIDIA GB10 Grace Blackwell run local LLMs?
Yes. With 128 GB (128 GB usable by a model) it runs 144 of the catalogued models at Q4_K_M with 8K context; the largest is Step 3.7 Flash 198B-A11B, needing about 119.7 GB.
How fast is the NVIDIA GB10 Grace Blackwell for AI inference?
It is estimated to run Llama 3.1 8B at 26–53 tok/s at Q4_K_M. Llama 3.3 70B is estimated at 3.4–7.1 tok/s. These are modelled estimates from memory bandwidth, not measurements; the methodology page shows the formula.
What LLMs can I run on 128 GB?
Among the best that fit: Qwen3.8-Flash-Next, Qwen 3.5 122B-A10B, Nemotron 3 Super 120B-A12B, Nemotron 3.5 Super VL 120B-A12B, Mistral Small 4 119B-A6.5B.