NVIDIA GeForce RTX 4090
NVIDIA · 24GB GDDR6X · Can run 20 models
| Manufacturer | NVIDIA |
| VRAM | 24 GB |
| Memory Type | GDDR6X |
| Architecture | Ada Lovelace |
| CUDA Cores | 16,384 |
| Tensor Cores | 512 |
| TDP | 450W |
| MSRP | $1,599 |
| Released | Oct 12, 2022 |
AI Notes
The RTX 4090 remains one of the best GPUs for local AI inference. Its 24GB of GDDR6X VRAM can run 13B models at full precision and 30B+ models with quantization. The massive tensor core count delivers class-leading inference throughput among consumer GPUs.
Compatible Models
| Model | Parameters | Best Quant | VRAM Used | Fit |
|---|---|---|---|---|
| Llama 3.2 1B | 1B | Q8_0 | 3 GB | Runs |
| Gemma 2 2B | 2B | Q8_0 | 4 GB | Runs |
| Llama 3.2 3B | 3B | Q8_0 | 5 GB | Runs |
| Phi-3 Mini 3.8B | 3.8B | Q8_0 | 5.8 GB | Runs |
| DeepSeek R1 7B | 7B | Q8_0 | 9 GB | Runs |
| Mistral 7B | 7B | Q8_0 | 9 GB | Runs |
| Qwen 2.5 7B | 7B | Q8_0 | 9 GB | Runs |
| Qwen 2.5 Coder 7B | 7B | Q8_0 | 9 GB | Runs |
| Llama 3.1 8B | 8B | Q8_0 | 10 GB | Runs |
| Gemma 2 9B | 9B | Q8_0 | 11 GB | Runs |
| DeepSeek R1 14B | 14B | Q4_K_M | 9.9 GB | Runs |
| Phi-4 14B | 14B | Q4_K_M | 9.9 GB | Runs |
| Qwen 2.5 14B | 14B | Q4_K_M | 9.9 GB | Runs |
| StarCoder2 15B | 15B | Q8_0 | 17 GB | Runs |
| Codestral 22B | 22B | Q4_K_M | 14.7 GB | Runs |
| Gemma 2 27B | 27B | Q4_K_M | 17.7 GB | Runs |
| DeepSeek R1 32B | 32B | Q4_K_M | 20.7 GB | Runs (tight) |
| Qwen 2.5 32B | 32B | Q4_K_M | 20.7 GB | Runs (tight) |
| Command R 35B | 35B | Q4_K_M | 22.5 GB | Runs (tight) |
| Mixtral 8x7B | 47B | Q4_K_M | 29.7 GB | CPU Offload |
5
model(s) are too large for this hardware.