Lenovo ThinkStation P3 Ultra (RTX PRO 4500)
NVIDIA RTX PRO 4500 Blackwell · AI Workstation
The Lenovo ThinkStation P3 Ultra (RTX PRO 4500) is Lenovo's workstation built on the NVIDIA RTX PRO 4500 Blackwell, and it gives a local model about 32GB of usable AI memory at 896 GB/s. The 32GB / 512GB configuration is $2,029.99 at Amazon, verified 5h ago. It scores 39.5 overall on the AIPC Score.
Price and memory shown are for the 32GB / 512GB configuration.
AIPC Score: 39.5
- AI memory33.3
- AI compute24
- Memory bandwidth49.8
- Model compatibility58.6
- GPU capability36.9
- NPU capability0
- Software support100
- Power efficiency7.1
- Expandability85
- Price/value19.7
Specifications
- AI Memory
- 32GB
- CPU
- Intel Core Ultra 9 285
- GPU
- RTX PRO 4500 Blackwell
- NPU
- —
- Memory bandwidth
- 896 GB/s
- Storage
- 512GB
- Configuration
- 32GB / 512GB
- Type
- Workstation
How we calculate AI memory: Dedicated VRAM on the graphics card. System RAM is listed separately as CPU offload capacity. Source →
Configurations
| Configuration | AI Memory | Memory bandwidth | Price | Retailer |
|---|---|---|---|---|
| 32GB / 512GB | 32GB | 896 GB/s | $2,029.99 | Amazon (third-party seller) |
| 64GB / 2TB | 32GB | 896 GB/s | $3,499 | Amazon (third-party seller) |
| 128GB / 2TB | 32GB | 896 GB/s | $5,998.99 | Amazon (third-party seller) |
What can this PC run?
| Model | Quantization | Fits | Mode | State | Speed |
|---|---|---|---|---|---|
| DeepSeek-R1-Distill-Qwen-7B | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| DeepSeek-R1-Distill-Qwen-7B | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| DeepSeek-R1-Distill-Qwen-14B | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| DeepSeek-R1-Distill-Qwen-14B | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| DeepSeek-R1-Distill-Qwen-32B | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Gemma 3 4B IT | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Gemma 3 4B IT | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Gemma 3 12B IT | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Gemma 3 12B IT | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Gemma 3 27B IT | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Gemma 3 27B IT | Q8 | Limited | Offload required | Calculated | Performance not yet benchmarked |
| Llama 3.2 3B Instruct | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Llama 3.2 3B Instruct | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Llama 3.1 8B Instruct | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Llama 3.1 8B Instruct | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Mistral Small 3 (Mistral-Small-24B-Instruct-2501) | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Mistral Small 3 (Mistral-Small-24B-Instruct-2501) | Q8 | Limited | Offload required | Calculated | Performance not yet benchmarked |
| Devstral Small 1.1 (Devstral-Small-2507) | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Devstral Small 1.1 (Devstral-Small-2507) | Q8 | Limited | Offload required | Calculated | Performance not yet benchmarked |
| Phi-4 | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Phi-4 | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Qwen3-4B | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Qwen3-4B | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Qwen3-8B | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Qwen3-8B | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Qwen3-14B | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Qwen3-14B | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Qwen3-30B-A3B | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Qwen3-30B-A3B | Q8 | Limited | Offload required | Calculated | Performance not yet benchmarked |
| Qwen3-32B | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| gpt-oss-20b | Q4 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| gpt-oss-20b | Q8 | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Whisper large-v3 | Published | Yes | Full memory | Calculated | Performance not yet benchmarked |
| FLUX.1 schnell | Published | Yes | Full memory | Calculated | Performance not yet benchmarked |
| FLUX.1 dev | Published | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Stable Diffusion XL Base 1.0 | Published | Yes | Full memory | Calculated | Performance not yet benchmarked |
| Stable Diffusion 3.5 Large | Published | Yes | Full memory | Calculated | Performance not yet benchmarked |
Does not fit (29)
- DeepSeek-R1-Distill-Qwen-32B Q8: needs 42.3GB
- DeepSeek-R1-Distill-Llama-70B Q4: needs 51.5GB
- DeepSeek-R1-Distill-Llama-70B Q8: needs 88.8GB
- DeepSeek-R1 Q4: needs 469.8GB
- DeepSeek-R1 Q8: needs 821.9GB
- DeepSeek-V3 Q4: needs 469.8GB
- DeepSeek-V3 Q8: needs 821.9GB
- GLM-5.2 Q4: needs 527GB
- GLM-5.2 Q8: needs 922.1GB
- Kimi-K2-Instruct Q4: needs 699.2GB
- Kimi-K2-Instruct Q8: needs 1223.9GB
- Llama 3.1 70B Instruct Q4: needs 50.8GB
- Llama 3.1 70B Instruct Q8: needs 87.5GB
- Llama 3.3 70B Instruct Q4: needs 50.8GB
- Llama 3.3 70B Instruct Q8: needs 87.5GB
- Llama 4 Scout 17B-16E Instruct Q4: needs 78GB
- Llama 4 Scout 17B-16E Instruct Q8: needs 135.2GB
- Llama 4 Maverick 17B-128E Instruct Q4: needs 280.9GB
- Llama 4 Maverick 17B-128E Instruct Q8: needs 490.7GB
- Mistral Medium 3.5 128B Q4: needs 91.2GB
- Mistral Medium 3.5 128B Q8: needs 158.4GB
- Qwen3-32B Q8: needs 42.1GB
- Qwen3-235B-A22B Q4: needs 165.8GB
- Qwen3-235B-A22B Q8: needs 289.1GB
- gpt-oss-120b Q4: needs 83.6GB
- gpt-oss-120b Q8: needs 145GB
- HunyuanVideo Published: needs 60GB
- Wan2.1 T2V 14B Published: needs 65.7GB
- Wan2.2 T2V A14B Published: needs 80GB
Alternatives
- Corsair AI Workstation 300
96GB · AI Score 39.5
No verified price
- Acer Veriton RA100
96GB · AI Score 39.8
No verified price
- HP Z2 Tower G1i (RTX PRO 4500)
32GB · AI Score 38.5
No verified price