NVIDIA RTX PRO 4000 Blackwell SFF for local LLMs
Written by Jakub Rusinowski · Last updated
With 24 GB of GDDR7 at 432 GB/s, the RTX PRO 4000 Blackwell SFF runs 137 catalogued models at Q4_K_M with 8K context. The largest that fits is Qwen 3 32B (~22.8 GB), and the top pick is Laguna XS 2.1 33B-A3B at 53–111 tok/s.
The low-profile, 70 W version of the RTX PRO 4000 Blackwell: the same 24 GB of GDDR7 with ECC, but on a narrower 192-bit bus at 432 GB/s instead of 672 GB/s, so decode is about a third slower than the full-size card. Powered from the slot, for compact workstations.
Models that run on the RTX PRO 4000 Blackwell SFF
Q4_K_M, 8K context, 24 GB usable. Ranked by quality and speed.
| Model | Memory · marker = 24 GB | VRAM | Speed |
|---|---|---|---|
| Laguna XS 2.1 33B-A3B Poolside Laguna XS 2.1 | 20.7 GB | 53–111 tok/s | |
| Granite 4.0 Small-H 32B-A9B IBM Granite 4.0 | 20.1 GB | 29–61 tok/s | |
| Nemotron-Cascade 2 30B-A3B Nemotron Cascade 2 | 19.9 GB | 50–104 tok/s | |
| Qwen 3 30B-A3B (MoE) Qwen 3 | 20 GB | 66–137 tok/s | |
| Qwen3-Coder 30B-A3B (MoE) Qwen3-Coder | 20 GB | 66–137 tok/s | |
| GLM-4.7-Flash 30B-A3B GLM-4.7 / GLM-Z1 | 18.9 GB | 54–112 tok/s | |
| Nemotron 3 Nano Omni 30B-A3B Nemotron 3 Nano Omni | 18.9 GB | 54–112 tok/s | |
| North Mini Code 1.0 30B-A3B North Mini Code | 18.9 GB | 54–112 tok/s | |
| Nemotron 3.5 Lightning 30B-A3B Nemotron 3.5 | 18.9 GB | 54–112 tok/s | |
| Granite 4.1 30B IBM Granite 4.1 | 18.9 GB | 12–24 tok/s | |
| Granite 4.2 30B IBM Granite 4.2 | 18.9 GB | 12–24 tok/s | |
| Muse Glimmer 30B Muse Glimmer | 18.7 GB | 12–24 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.
or compare on Vast.ai from $0.35/hr (typical low · varies)
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 2026-11-11.
- Memory
- 24 GB GDDR7
- Memory bandwidth
- 432 GB/s
- Memory bus
- 192-bit
- Architecture
- Blackwell GB203
- Series
- RTX PRO Blackwell
- Board power
- 70 W
- Release year
- 2025
- Compute backends
- CUDA
- Usable for models
- 24 GB
Similar GPUs
Frequently asked questions
Can the NVIDIA RTX PRO 4000 Blackwell SFF run local LLMs?
Yes. With 24 GB (24 GB usable by a model) it runs 137 of the catalogued models at Q4_K_M with 8K context; the largest is Qwen 3 32B, needing about 22.8 GB.
How fast is the NVIDIA RTX PRO 4000 Blackwell SFF for AI inference?
It is estimated to run Llama 3.1 8B at 39–80 tok/s at Q4_K_M. Llama 3.3 70B does not fit: it needs about 44 GB against 24 GB usable. These are modelled estimates from memory bandwidth, not measurements; the methodology page shows the formula.
What LLMs can I run on 24 GB?
Among the best that fit: Laguna XS 2.1 33B-A3B, Granite 4.0 Small-H 32B-A9B, Nemotron-Cascade 2 30B-A3B, Qwen 3 30B-A3B (MoE), Qwen3-Coder 30B-A3B (MoE).