NVIDIA · Blackwell · 2025
NVIDIA RTX PRO 5000 Blackwell 72GB for local AI
72GB on one 300W card fits gpt-oss-120b whole, for a little over half the price of an RTX PRO 6000.
Updated · Speeds marked ~ are estimates; seemethodology
- VRAM
- 72GB
- Bandwidth
- 1,344 GB/s
- Power
- 300W
- Runs well
- 12 models
RTX PRO 5000 72GB
Typically about $9,100
Can the RTX PRO 5000 72GB run LLMs?
The RTX PRO 5000 72GB has 72GB of GDDR7 ECC and 1,344 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 Q3_K_M (62GB with 8K context, ~99 tokens/sec). Our top all-round pick for it is Qwen3.8 27B (Q8_0, ~35 tokens/sec).
Recommendations
Best models for the RTX PRO 5000 72GB
Best overall · also coding, reasoning, vision
Qwen3.8 27BQ8_0 · 32GB · ~35 tokens/sec
Runs wellRunner-up
Qwen3.6 35B-A3BQ8_0 · 41GB · ~130 tokens/sec
Runs wellFastest good model
gpt-oss-20bMXFP4 · 15GB · ~160 tokens/sec
Runs wellLargest model that fits
Qwen3.5 122B-A10BQ3_K_M · 62GB · ~99 tokens/sec
Runs well
Compatibility
What runs on the RTX PRO 5000 72GB
Every current model in our catalog, at the best quantization that fits in 72GB with 8K tokens of context. "Partial offload" means part of the model runs from system RAM (we assume 64GB), which is much slower.
| Model | Parameters | Best quant | Memory | Speed | Verdict |
|---|---|---|---|---|---|
| Gemma 4 E4B | 8B | Q8_0 | 10 GB | ~100 tok/s | Runs well |
| Qwen3.5 9B | 9B | Q8_0 | 12 GB | ~90 tok/s | Runs well |
| Phi-4 | 14.7B | Q8_0 | 18 GB | ~65 tok/s | Runs well |
| gpt-oss-20b | 20.9B (3.6B active) | MXFP4 | 15 GB | 215 tok/s | Runs well |
| Gemma 4 26B A4B | 26B (3.8B active) | Q8_0 | 30 GB | ~115 tok/s | Runs well |
| Qwen3.8 27B | 27B | Q8_0 | 32 GB | ~35 tok/s | Runs well |
| Gemma 4 31B | 31B | Q8_0 | 36 GB | ~32 tok/s | Runs well |
| Qwen3.6 35B-A3B | 35B (3B active) | Q8_0 | 41 GB | ~130 tok/s | Runs well |
| GLM-4.5-Air | 106B (12B active) | Q4_K_M | 68 GB | ~73 tok/s | Fits, tight |
| Llama 4 Scout | 109B (17B active) | Q3_K_M | 56 GB | ~66 tok/s | Runs well |
| gpt-oss-120b | 117B (5.1B active) | MXFP4 | 68 GB | ~145 tok/s | Fits, tight |
| Qwen3.5 122B-A10B | 122B (10B active) | Q3_K_M | 62 GB | ~99 tok/s | Runs well |
| DeepSeek V4 Flash | 284B (13B active) | — | — | — | Won't fitWhat it needs |
| GLM-5.3-Flash | 320B (18B active) | — | — | — | Won't fitWhat it needs |
| GLM-5.3 | 744B (40B active) | — | — | — | Won't fitWhat it needs |
Performance
RTX PRO 5000 72GB vs. alternatives on Qwen3.8 27B
Generation speed in tokens per second for a single user. Token generation is limited mostly by memory bandwidth, which is why bandwidth predicts these results so well.
RTX 5090
Q6_K · 32GB
~59tok/s
RTX PRO 6000
Q8_0 · 96GB
~46tok/s
RTX PRO 5000 72GB
Q8_0 · 72GB
~35tok/s
RTX 6000 Ada
Q8_0 · 48GB
~25tok/s
Measured
RTX PRO 5000 72GB benchmark results
| Model | Quant | Runtime | Generation | Prompt | Source |
|---|---|---|---|---|---|
| gpt-oss-20b | MXFP4 | llama.cpp (CUDA) | 215 tok/s | 10340 tok/s | Hardware Corner: RTX Pro 5000 Blackwell LLM Performance · March 2026 |
| Qwen3 30B-A3B (2507) | Q4_K_XL | llama.cpp (CUDA) | 190 tok/s | 5265 tok/s | Hardware Corner: RTX Pro 5000 Blackwell LLM Performance · March 2026 |
| Qwen3 8B | Q4_K_XL | llama.cpp (CUDA) | 160 tok/s | 8175 tok/s | Hardware Corner: RTX Pro 5000 Blackwell LLM Performance · March 2026 |
| Qwen3 14B | Q4_K_XL | llama.cpp (CUDA) | 99 tok/s | 4515 tok/s | Hardware Corner: RTX Pro 5000 Blackwell LLM Performance · March 2026 |
| Qwen3 32B | Q4_K_XL | llama.cpp (CUDA) | 47 tok/s | 2030 tok/s | Hardware Corner: RTX Pro 5000 Blackwell LLM Performance · March 2026 |
| Llama 3.3 70B Instruct | Q4_K_XL | llama.cpp (CUDA) | 24 tok/s | 1025 tok/s | Hardware Corner: RTX Pro 5000 Blackwell LLM Performance · March 2026 |
Our take
Should you buy the RTX PRO 5000 72GB for AI?
The 72GB RTX PRO 5000 Blackwell became generally available in December 2025 with 72GB of ECC GDDR7 on a 384-bit bus for 1,344 GB/s. That is enough to hold gpt-oss-120b entirely on one card, and its 300W rating makes multi-GPU workstations practical. It sits between the 48GB RTX PRO 5000 and the 96GB RTX PRO 6000 in NVIDIA's lineup. Street prices were about $9,100 to $9,575 in August 2026.
What we like
- 72GB on a single card
- 1,344 GB/s GDDR7 bandwidth
- 300W design suits multi-GPU towers
- ECC memory and professional drivers
Watch out for
- Costs $9,100 or more
- Less memory than the RTX PRO 6000 for the largest models
Best for: Running gpt-oss-120b, or 70B models at 6 bit, on one card · Workstations with two to four GPUs
Specifications
RTX PRO 5000 72GB specs that matter for AI
- Memory
- 72GB GDDR7 ECC
- Memory bandwidth
- 1,344 GB/s
- Memory bus
- 384-bit
- Architecture
- Blackwell
- Power
- 300W
- Interface
- PCIe 5.0 x16
- Released
- 2025
- Street price (new)
- ~$9,100 (2026-10)
- Software
- CUDA, Vulkan
Sources: NVIDIA blog: RTX PRO 5000 72GB Blackwell now generally available, RunAIHome RTX PRO 5000 72GB pricing and fit analysis
Cross-shop
Alternatives to the RTX PRO 5000 72GB
Frequently asked questions
How much VRAM does the RTX PRO 5000 72GB have?
The NVIDIA RTX PRO 5000 Blackwell 72GB has 72GB of GDDR7 ECC with 1,344 GB/s of memory bandwidth on a 384-bit bus.
What is the largest LLM the RTX PRO 5000 72GB can run?
Fully in GPU memory, the largest model in our catalog that runs comfortably is Qwen3.5 122B-A10B (122B (10B active)) at Q3_K_M, using about 62GB with 8K context at ~99 tokens/sec.
Can the RTX PRO 5000 72GB run Llama 3.3 70B?
Yes. At Q6_K it needs about 63GB with 8K context and runs at ~18 tokens/sec.
What is the best LLM to run on the RTX PRO 5000 72GB?
Our top pick is Qwen3.8 27B at Q8_0, which fits in 72GB with room for 8K of context and runs at ~35 tokens/sec. See our full list of the best models for the RTX PRO 5000 72GB for coding, reasoning, and vision picks.
Is the RTX PRO 5000 72GB good for local AI in 2026?
72GB on one 300W card fits gpt-oss-120b whole, for a little over half the price of an RTX PRO 6000.