RTX PRO 4000 Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
The RTX PRO 4000 Blackwell is NVIDIA's 2025 single-slot 24GB, 145W card, running 30B-class MoE at about 100 tokens/s. The 70W SFF is a third slower.
eBay
RTX PRO 4000 price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Full RTX PRO 4000 specifications
- Memory24 GB · GDDR7 ECC
- Bandwidth672 GB/s
- FP16 / BF16161.3 TFLOPS
- FP8 / INT8322.5 / 322.5 TOPS
- Bus width
- 192-bit
- Architecture
- Blackwell (GB203)
- Cores
- 8,960
- Tensor cores
- 280
- Ecosystem
- CUDA
- Interface
- PCIe 5.0 x16
- NVLink
- Not supported
- Power
- 145 W
- Slots
- 1
- Length
- 241 mm
- Type
- Workstation
- Launch date
- Mar 18, 2025
Best RTX PRO 4000 deals on eBay
Hand-picked deals from trusted eBay sellers. Shop with confidence.
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
UsedNVIDIA RTX PRO 4000 Blackwell 24GB - Graphics Card - Very GoodTop Rated$2,660Buy on eBay
UsedHP RTX PRO 4000 Blackwell 24GB RAM - Graphics Card - Excellent ConditionTop Rated$2,730Buy on eBay
UsedDell NVIDIA RTX PRO 4000 Blackwell 24GB Graphics Card D2JWW VCNRTXPRO4000B-PB$2,900+$20 shippingBuy on eBay
UsedNVIDIA RTX PRO 4000 BLACKWELL GPU 24GB GRAPHICS VIDEO CARD 900-5G147-2570-000$2,950Buy on eBay
NewDELL NVIDIA RTX PRO 4000 Blackwell 24GB GDDR7 PCIe 5.0 OEM Graphics Card New$2,995Buy on eBay
UsedDell NVIDIA RTX PRO 4000 Blackwell 24GB GDDR7 PCIe 5.0 Graphics Card D2JWW 0D2JWTop Rated$2,999Buy on eBay
What LLMs can a single RTX PRO 4000 run?
A single PRO 4000 fits up to 35B at 4-bit; context is the squeeze, Gemma 4 31B only about 66K. If you feed it long documents, drop to 2-bit.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 E4B8B | 3.8 GB | Runs | 112 | 128K |
 Qwen3.5 9B9.7B | 4.1 GB | Runs | 104 | 256K |
| Gemma 4 12B12B | 4.7 GB | Runs | 91.2 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 123 | 256K |
| Qwen3.6 27B27.8B | 11.8 GB | Runs | 40.3 | 192K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 47.8 | 221K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 134.6 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 38.8 | 143K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 145.9 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 5 GB | Runs | 88.9 | 128K |
| Qwen3.5 9B9.7B | 5.7 GB | Runs | 78.6 | 256K |
| Gemma 4 12B12B | 7.1 GB | Runs | 64 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 99.9 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 29 | 117K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 29.5 | 122K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 115.3 | 115K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 26.1 | 66K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 118.5 | 123K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 8.2 GB | Runs | 57.3 | 128K |
| Qwen3.5 9B9.7B | 9.5 GB | Runs | 49.8 | 256K |
| Gemma 4 12B12B | 12.7 GB | Runs | 37.7 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Needs RAM offload | 25.3 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Too big | — | — |
| Qwen3.8 27B27.8B | 29 GB | Too big | — | — |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Needs RAM offload | 22.4 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Too big | — | — |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Needs RAM offload | 33.4 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Your case has room for only a single-slot card, or you want several in one workstation
Single slot, 145W, blower cooling
- Mainly running MoE models with few active parameters
Qwen3.6 35B A3B and GLM 4.7 Flash at 4-bit both run above 115 tokens/s
- You want a new card with ECC memory and FP8 and FP4 support
Not a good fit
- You need long context
Gemma 4 31B at 4-bit only reaches about 66K, and Qwen3.8 27B about 122K
- Generation speed is your priority
Qwen3.8 27B at 4-bit runs at about 30 tokens/s, versus about 40 on a used 3090
- Running 27B models at 8-bit
Qwen3.8 27B needs 29 GB at 8-bit, which doesn't fit
Before you buy
One 8-pin cable is enough
The card has one 16-pin power connector and draws 145W at full load. Per NVIDIA's power guide, an adapter from one PCIe 8-pin to 16-pin delivers up to 150W, which is just enough, so a power supply with only 8-pin cables can run it.
Don't confuse it with the SFF version
There is also an RTX PRO 4000 Blackwell SFF Edition, likewise 24GB, but it draws only 70W and is a half-height dual-slot card with about two-thirds the bandwidth of the standard version, so it runs models a third slower. If a listing just says “PRO 4000 24G”, ask whether it is the SFF.
Requires CUDA 12.8 or later
This generation is supported only from CUDA 12.8 onward, and PyTorch has matching builds only from 2.7. Old Docker images and older versions of vLLM won't run on this card, so upgrade the driver and frameworks first.
No MIG
Among this generation's RTX PRO workstation cards, only the 5000 and above support MIG (splitting one card into several). If you want to share one card among different people or services, this card can't do it.
Compared with similar GPUs
RTX PRO 4000 vs GeForce RTX 3090, Arc Pro B60, and RTX PRO 4500: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | RTX PRO 4000 | GeForce RTX 3090 |  Arc Pro B60 | RTX PRO 4500 |
|---|---|---|---|---|
| Memory | 24 GB | 24 GBSame | 24 GBSame | 32 GB33% more |
| Bandwidth | 672 GB/s | 936 GB/s39% higher | 456 GB/s32% lower | 896 GB/s33% higher |
| BF16 / FP16 compute | 161 TFLOPS | 142 TFLOPS12% lower | 99 TFLOPS39% lower | 211 TFLOPS31% higher |
| INT8 compute | 323 TOPS | 285 TOPS12% lower | 197 TOPS39% lower | 422 TOPS31% higher |
| Qwen3.8 27B speed (2-bit) | 47.8 t/s | 64.8 t/s36% faster | 33.1 t/s31% slower | 62.3 t/s30% faster |
| Largest dense model (4-bit) | 37B | 37BSame | 37BSame | 51B38% more |
| Power | 145 W | 350 W141% higher | 200 W38% higher | 200 W38% higher |
| GPU interconnect | Not supported | NVLink 112.5 GB/s | — | Not supported |
| Interface | PCIe 5.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x8 | PCIe 5.0 x16 |
| Released | 2025-03 | 2020-09 | 2025-08 | 2025-03 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
RTX PRO 4000 rental price per hour
On-demand rental of the RTX PRO 4000 starts at $0.34 per hour (Vast.ai), about $82/month at 8 hours a day. Buying one (eBay median $2,988) pays off in about 6.6 years, 3.3 years or 1.1 years at 4, 8 or 24 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
 Vast.aiRentUpdated 48 min. ago | $0.34/hr | $41/mo | $82/mo | $246/mo | Rent |
 RunPodRentUpdated 48 min. ago | $0.57/hr | $69/mo | $139/mo | $416/mo | Rent |
 eBayUpdated 6 hr. ago | $2,988 | Payback ~6.6 yrsPower $3/mo | Payback ~3.3 yrsPower $6/mo | Payback ~1.1 yrsPower $19/mo | Buy |
Break-even uses the US residential average ($0.18/kWh) and 145 W rated power, excluding resale value and the rest of the build.
Rental links are referral links: we may receive a commission or credit if you sign up through them. As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Amazon.jp
Xianyu