GPUs · NVIDIA
Can the RTX 4090 run local LLMs?
With 24 GB of VRAM (about 21.6 GB usable) and 1008 GB/s of memory bandwidth, the RTX 4090 runs 28 of 39 popular open models fully on the GPU at 8K context. The largest is Qwen3 30B-A3B (MoE) at Q4_K_M, at roughly 130 tokens per second.
Best picks for the RTX 4090
Every model on the RTX 4090
Results at 8K context, picking the highest quality quantization that fits. "Offload" means part of the model runs from system RAM (we assume 32 GB), which works but is much slower.
| Model | Status | Quant | Memory | Speed |
|---|---|---|---|---|
| Qwen3 30B-A3B (MoE) | Runs well | Q4_K_M | 21 GB | ~130 tok/s |
| Qwen3-Coder 30B-A3B | Runs well | Q4_K_M | 21 GB | ~130 tok/s |
| Qwen3.6 27B | Runs well | Q4_K_M | 19 GB | ~40 tok/s |
| Gemma 4 26B-A4B (MoE) | Runs well | Q4_K_M | 19 GB | ~120 tok/s |
| Mistral Small 3 24B | Runs well | Q5_K_M | 21 GB | ~38 tok/s |
| Devstral Small 24B | Runs well | Q5_K_M | 20 GB | ~38 tok/s |
| Codestral 22B | Runs well | Q5_K_M | 19 GB | ~42 tok/s |
| gpt-oss 20B | Runs well | Q6_K | 19 GB | ~97 tok/s |
| Qwen3 14B | Runs well | Q8_0 | 19 GB | ~42 tok/s |
| DeepSeek-R1 Distill 14B | Runs well | Q8_0 | 19 GB | ~42 tok/s |
| Phi-4 14B | Runs well | Q8_0 | 18 GB | ~42 tok/s |
| Gemma 3 12B | Runs well | Q8_0 | 17 GB | ~51 tok/s |
| Mistral Nemo 12B | Runs well | Q8_0 | 16 GB | ~51 tok/s |
| Gemma 4 12B | Runs well | Q8_0 | 17 GB | ~52 tok/s |
| Qwen3.5 9B | Runs well | Q8_0 | 12 GB | ~66 tok/s |
| Qwen3 8B | Runs well | Q8_0 | 11 GB | ~75 tok/s |
| Llama 3.1 8B | Runs well | Q8_0 | 10 GB | ~77 tok/s |
| DeepSeek-R1 Distill 8B | Runs well | Q8_0 | 10 GB | ~77 tok/s |
| Qwen2.5-Coder 7B | Runs well | Q8_0 | 9.8 GB | ~81 tok/s |
| Mistral 7B v0.3 | Runs well | Q8_0 | 9.4 GB | ~86 tok/s |
| Qwen3.5 4B | Runs well | Q8_0 | 6.4 GB | ~130 tok/s |
| Gemma 3 4B | Runs well | Q8_0 | 6.9 GB | ~140 tok/s |
| Qwen3 4B | Runs well | Q8_0 | 5.8 GB | ~150 tok/s |
| Phi-4 Mini 3.8B | Runs well | Q8_0 | 5.2 GB | ~160 tok/s |
| Llama 3.2 3B | Runs well | Q8_0 | 4.6 GB | ~190 tok/s |
| Llama 3.2 1B | Runs well | Q8_0 | 1.6 GB | ~520 tok/s |
| Gemma 3 1B | Runs well | Q8_0 | 1.4 GB | ~620 tok/s |
| Qwen3 0.6B | Runs well | Q8_0 | 0.9 GB | ~1030 tok/s |
| Gemma 3 27B | Offload | Q4_K_M | 22 GB | ~5.9 tok/s |
| Qwen3 32B | Offload | Q4_K_M | 23 GB | ~4.8 tok/s |
| DeepSeek-R1 Distill 32B | Offload | Q4_K_M | 23 GB | ~4.8 tok/s |
| Qwen2.5-Coder 32B | Offload | Q4_K_M | 23 GB | ~4.8 tok/s |
| Qwen3.6 35B-A3B (MoE) | Offload | Q4_K_M | 23 GB | ~20 tok/s |
| Gemma 4 31B | Offload | Q4_K_M | 25 GB | ~4.9 tok/s |
| Llama 3.3 70B | Offload | Q3_K_M | 40 GB | ~2.0 tok/s |
| DeepSeek-R1 Distill 70B | Offload | Q3_K_M | 40 GB | ~2.0 tok/s |
| Llama 4 Scout 109B-A17B | Won’t fit | — | 73 GB | — |
| gpt-oss 120B | Won’t fit | — | 77 GB | — |
| Qwen3 235B-A22B (MoE) | Won’t fit | — | 155 GB | — |
What this means in practice
The RTX 4090 is comfortable with models up to about 30B parameters. Qwen3.6 35B-A3B (MoE) is the most useful step up beyond that: it needs about 23 GB, so part of it runs from system RAM, but as a mixture-of-experts model it stays reasonably quick at around 20 tokens per second. Models from Llama 4 Scout 109B-A17B upward need 73 GB or more and are out of reach. NVIDIA cards have the broadest support: Ollama, LM Studio and llama.cpp all use CUDA out of the box.
The next memory tier up is 32 GB: RTX 5090. Other 24 GB cards: RTX 3090 / 3090 Ti, RX 7900 XTX.
Compare the RTX 4090
Questions
What is the biggest model the RTX 4090 can run?
At 8K context the largest model that fits entirely in its 24 GB of VRAM is Qwen3 30B-A3B (MoE) at Q4_K_M, at roughly 130 tokens per second.
Can the RTX 4090 run 70B models like Llama 3.3 70B?
Only with heavy CPU offload. It needs about 40 GB at Q4, far more than 24 GB, so speed drops to a few tokens per second. A 48 GB setup or a Mac with 64 GB or more is the practical route.
What is the best model for coding on the RTX 4090?
Qwen3 30B-A3B (MoE) at Q4_K_M fits fully and runs at about 130 tokens per second. Pair it with an editor extension like Continue pointed at Ollama.
Estimates based on memory size and bandwidth; real speed varies with drivers, software and context length. See how we calculate VRAM.