Compare · GPUs
RTX 5080 vs RTX 4090 for local LLMs
It runs 28 of 39 models fully in memory, against 21 for the RTX 5080, thanks to 24 GB of VRAM versus 16 GB. The RTX 5080 is about 6% faster on the models both can hold, so it only wins if you stick to small models.
It runs 28 of 39 models fully in memory, against 21 for the RTX 5080, thanks to 24 GB of VRAM versus 16 GB. The RTX 5080 is about 6% faster on the models both can hold, so it only wins if you stick to small models.
Every model at 8K context with the best quantization that fits. Offload assumes 32 GB of system RAM. Green marks the better result. Rows where the two differ come first.
| Model | RTX 5080 | RTX 4090 |
|---|---|---|
| Codestral 22B | Offload~6.6 tok/sQ4_K_M | Runs~42 tok/sQ5_K_M |
| Mistral Small 3 24B | Offload~5.8 tok/sQ4_K_M | Runs~39 tok/sQ5_K_M |
| Devstral Small 24B | Offload~5.9 tok/sQ4_K_M | Runs~39 tok/sQ5_K_M |
| Gemma 4 26B-A4B (MoE) | Offload~16 tok/sQ4_K_M | Runs~120 tok/sQ4_K_M |
| Qwen3.6 27B | Offload~4.9 tok/sQ4_K_M | Runs~40 tok/sQ4_K_M |
| Qwen3 30B-A3B (MoE) | Offload~17 tok/sQ4_K_M | Runs~130 tok/sQ4_K_M |
| Qwen3-Coder 30B-A3B | Offload~17 tok/sQ4_K_M | Runs~130 tok/sQ4_K_M |
| Llama 3.3 70B | Won’t fit | Offload~2.0 tok/sQ3_K_M |
| DeepSeek-R1 Distill 70B | Won’t fit | Offload~2.0 tok/sQ3_K_M |
| Qwen3 0.6B | Runs~980 tok/sQ8_0 | Runs~1030 tok/sQ8_0 |
| Gemma 3 1B | Runs~590 tok/sQ8_0 | Runs~620 tok/sQ8_0 |
| Llama 3.2 1B | Runs~490 tok/sQ8_0 | Runs~520 tok/sQ8_0 |
| Llama 3.2 3B | Runs~180 tok/sQ8_0 | Runs~190 tok/sQ8_0 |
| Phi-4 Mini 3.8B | Runs~150 tok/sQ8_0 | Runs~160 tok/sQ8_0 |
| Qwen3 4B | Runs~150 tok/sQ8_0 | Runs~150 tok/sQ8_0 |
| Gemma 3 4B | Runs~140 tok/sQ8_0 | Runs~140 tok/sQ8_0 |
| Qwen3.5 4B | Runs~130 tok/sQ8_0 | Runs~130 tok/sQ8_0 |
| Mistral 7B v0.3 | Runs~82 tok/sQ8_0 | Runs~86 tok/sQ8_0 |
| Qwen2.5-Coder 7B | Runs~78 tok/sQ8_0 | Runs~81 tok/sQ8_0 |
| Llama 3.1 8B | Runs~74 tok/sQ8_0 | Runs~77 tok/sQ8_0 |
| DeepSeek-R1 Distill 8B | Runs~74 tok/sQ8_0 | Runs~77 tok/sQ8_0 |
| Qwen3 8B | Runs~72 tok/sQ8_0 | Runs~75 tok/sQ8_0 |
| Qwen3.5 9B | Runs~63 tok/sQ8_0 | Runs~66 tok/sQ8_0 |
| Gemma 4 12B | Runs~63 tok/sQ6_K | Runs~52 tok/sQ8_0 |
| Gemma 3 12B | Runs~62 tok/sQ6_K | Runs~51 tok/sQ8_0 |
| Mistral Nemo 12B | Runs~62 tok/sQ6_K | Runs~51 tok/sQ8_0 |
| Phi-4 14B | Runs~60 tok/sQ5_K_M | Runs~42 tok/sQ8_0 |
| Qwen3 14B | Runs~59 tok/sQ5_K_M | Runs~42 tok/sQ8_0 |
| DeepSeek-R1 Distill 14B | Runs~59 tok/sQ5_K_M | Runs~42 tok/sQ8_0 |
| gpt-oss 20B | Runs~130 tok/sQ4_K_M | Runs~97 tok/sQ6_K |
| Gemma 3 27B | Offload~4.5 tok/sQ4_K_M | Offload~5.9 tok/sQ4_K_M |
| Gemma 4 31B | Offload~3.8 tok/sQ4_K_M | Offload~4.9 tok/sQ4_K_M |
| Qwen3 32B | Offload~3.7 tok/sQ4_K_M | Offload~4.8 tok/sQ4_K_M |
| DeepSeek-R1 Distill 32B | Offload~3.7 tok/sQ4_K_M | Offload~4.8 tok/sQ4_K_M |
| Qwen2.5-Coder 32B | Offload~3.7 tok/sQ4_K_M | Offload~4.8 tok/sQ4_K_M |
| Qwen3.6 35B-A3B (MoE) | Offload~17 tok/sQ4_K_M | Offload~20 tok/sQ4_K_M |
| Llama 4 Scout 109B-A17B | Won’t fit | Won’t fit |
| gpt-oss 120B | Won’t fit | Won’t fit |
| Qwen3 235B-A22B (MoE) | Won’t fit | Won’t fit |
Which should you buy?
The RTX 4090 is the better card for local AI. It runs 28 of 39 models fully in memory, against 21 for the RTX 5080, thanks to 24 GB of VRAM versus 16 GB. The RTX 5080 is about 6% faster on the models both can hold, so it only wins if you stick to small models.
Only the RTX 4090 fits Codestral 22B, Mistral Small 3 24B, Devstral Small 24B, Gemma 4 26B-A4B (MoE), Qwen3.6 27B, Qwen3 30B-A3B (MoE) and Qwen3-Coder 30B-A3B fully in memory.
See every result on its own page: RTX 5080 and RTX 4090.
Questions
Is the RTX 5080 or the RTX 4090 better for running LLMs?
The RTX 4090 is the better card for local AI. It runs 28 of 39 models fully in memory, against 21 for the RTX 5080, thanks to 24 GB of VRAM versus 16 GB. The RTX 5080 is about 6% faster on the models both can hold, so it only wins if you stick to small models.
Which models can the RTX 4090 run that the RTX 5080 can't?
At 8K context, only the RTX 4090 holds Codestral 22B, Mistral Small 3 24B, Devstral Small 24B, Gemma 4 26B-A4B (MoE), Qwen3.6 27B and Qwen3 30B-A3B (MoE) and 1 more fully in VRAM.
Does the RTX 5080 or RTX 4090 need anything else to run local AI?
Only the software. Install Ollama or LM Studio, pick a model that fits in VRAM, and it runs on the GPU. 32 GB of system RAM helps when a model spills over, but that path is much slower.
More comparisons
Estimates from memory size and bandwidth at 8K context. Real speed varies with software, drivers and settings.