RTX PRO 4000 Blackwell

The RTX PRO 4000 Blackwell is an NVIDIA GPU built on the Blackwell 2.0 architecture, released 2025-03-18. For running AI locally, the numbers that matter are its 24 GB of GDDR7 and 672 GB/s of memory bandwidth. VRAM decides which models fit at all; bandwidth sets how fast they generate text.

What you can run on 24 GB

At Q4_K_M quantization (the usual local default), 24 GB holds models up to roughly 38B parameters, leaving headroom for context. On this card you can run, among others:

• Qwen3.5 35B A3B - 36B parameters
• Qwen3.6 35B A3B - 36B parameters
• dolphin-2.9.1-yi-1.5-34b - 34.4B parameters
• EXAONE 4.5 33B - 34.4B parameters
• Qwen2.5-32B-Instruct - 32.8B parameters

Larger models need a higher-VRAM card, a second GPU, or CPU offload (which is much slower). Check any specific model with the VRAM calculator, or see the full picture on what can I run.

Local LLM speed (LLaMA 3, llama.cpp)

Single-stream token-generation throughput - estimated from memory bandwidth:

Model (quant)	Speed on RTX PRO 4000 Blackwell
Llama 3 8B (Q4_K_M)	78.6 tok/s
Llama 3 8B (F16)	35.7 tok/s
Llama 3 70B (Q4_K_M)	✗ won't fit

Because decode is memory-bandwidth bound, the 672 GB/s figure is the best single predictor of chat speed on this card. Estimates are calibrated against measured RTX-40-series cards and are typically within ~15%.

Memory and power

• VRAM: 24 GB GDDR7 (192-bit bus)
• Bandwidth: 672 GB/s
• TDP: 140 W - a 300 W+ power supply is recommended
• Process: 5 nm
• Interface: PCIe 5.0 x16

Quantization and context

Quantization trades a little quality for a lot of VRAM. On 24 GB you can fit roughly a 38B model at Q4_K_M, about a 20B model at the higher-quality Q8, or a smaller model at full FP16. Longer context windows also consume VRAM (the KV cache grows with context length), so leave a few GB of headroom if you plan to use large prompts or many concurrent requests. For most chat and coding use, Q4_K_M on this card is the sweet spot between speed, quality, and the 24 GB budget.

How it compares

Similar cards for local AI, by VRAM and 8B-Q4 speed:

GPU	VRAM	Bandwidth	Llama 3 8B Q4
RTX PRO 4000 Blackwell	24 GB	672 GB/s	78.6 tok/s
AMD Radeon RX 7900 XTX	24 GB	960 GB/s	112.3 tok/s
Arc Pro B60	24 GB	456 GB/s	53.4 tok/s
NVIDIA Tesla P40	24 GB	347 GB/s	40.6 tok/s

Bottom line

The RTX PRO 4000 Blackwell is best for llm-inference, image-gen, comfyui. With 24 GB+ it comfortably handles most open models, including 30B-class at Q4. If you need more, compare with AMD Radeon RX 7900 XTX and Arc Pro B60.

Sources

• Specifications: RightNow GPU Database (TechPowerUp data)
• Benchmarks: GPU-Benchmarks-on-LLM-Inference (basis for the bandwidth estimate)

Specs and benchmarks last checked 2026-06-08. Verify current pricing before buying.