Compare · GPUs
RTX 3090 / 3090 Ti vs RTX 4080 / 4080 Super for local LLMs
It runs 28 of 39 models fully in memory, against 21 for the RTX 4080 / 4080 Super, thanks to 24 GB of VRAM versus 16 GB. It is also about 14% faster on the models both can run.
It runs 28 of 39 models fully in memory, against 21 for the RTX 4080 / 4080 Super, thanks to 24 GB of VRAM versus 16 GB. It is also about 14% faster on the models both can run.
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 3090 / 3090 Ti | RTX 4080 / 4080 Super |
|---|---|---|
| Codestral 22B | Runs~39 tok/sQ5_K_M | Offload~5.2 tok/sQ4_K_M |
| Mistral Small 3 24B | Runs~36 tok/sQ5_K_M | Offload~4.6 tok/sQ4_K_M |
| Devstral Small 24B | Runs~36 tok/sQ5_K_M | Offload~4.7 tok/sQ4_K_M |
| Gemma 4 26B-A4B (MoE) | Runs~110 tok/sQ4_K_M | Offload~13 tok/sQ4_K_M |
| Qwen3.6 27B | Runs~38 tok/sQ4_K_M | Offload~3.9 tok/sQ4_K_M |
| Qwen3 30B-A3B (MoE) | Runs~120 tok/sQ4_K_M | Offload~14 tok/sQ4_K_M |
| Qwen3-Coder 30B-A3B | Runs~120 tok/sQ4_K_M | Offload~14 tok/sQ4_K_M |
| Llama 3.3 70B | Offload~1.9 tok/sQ3_K_M | Won’t fit |
| DeepSeek-R1 Distill 70B | Offload~1.9 tok/sQ3_K_M | Won’t fit |
| Qwen3 0.6B | Runs~960 tok/sQ8_0 | Runs~750 tok/sQ8_0 |
| Gemma 3 1B | Runs~570 tok/sQ8_0 | Runs~450 tok/sQ8_0 |
| Llama 3.2 1B | Runs~480 tok/sQ8_0 | Runs~380 tok/sQ8_0 |
| Llama 3.2 3B | Runs~180 tok/sQ8_0 | Runs~140 tok/sQ8_0 |
| Phi-4 Mini 3.8B | Runs~150 tok/sQ8_0 | Runs~120 tok/sQ8_0 |
| Qwen3 4B | Runs~140 tok/sQ8_0 | Runs~110 tok/sQ8_0 |
| Gemma 3 4B | Runs~130 tok/sQ8_0 | Runs~110 tok/sQ8_0 |
| Qwen3.5 4B | Runs~120 tok/sQ8_0 | Runs~96 tok/sQ8_0 |
| Mistral 7B v0.3 | Runs~80 tok/sQ8_0 | Runs~63 tok/sQ8_0 |
| Qwen2.5-Coder 7B | Runs~76 tok/sQ8_0 | Runs~59 tok/sQ8_0 |
| Llama 3.1 8B | Runs~72 tok/sQ8_0 | Runs~56 tok/sQ8_0 |
| DeepSeek-R1 Distill 8B | Runs~72 tok/sQ8_0 | Runs~56 tok/sQ8_0 |
| Qwen3 8B | Runs~70 tok/sQ8_0 | Runs~55 tok/sQ8_0 |
| Qwen3.5 9B | Runs~61 tok/sQ8_0 | Runs~48 tok/sQ8_0 |
| Gemma 4 12B | Runs~48 tok/sQ8_0 | Runs~49 tok/sQ6_K |
| Gemma 3 12B | Runs~47 tok/sQ8_0 | Runs~48 tok/sQ6_K |
| Mistral Nemo 12B | Runs~47 tok/sQ8_0 | Runs~48 tok/sQ6_K |
| Phi-4 14B | Runs~39 tok/sQ8_0 | Runs~46 tok/sQ5_K_M |
| Qwen3 14B | Runs~39 tok/sQ8_0 | Runs~46 tok/sQ5_K_M |
| DeepSeek-R1 Distill 14B | Runs~39 tok/sQ8_0 | Runs~46 tok/sQ5_K_M |
| gpt-oss 20B | Runs~90 tok/sQ6_K | Runs~98 tok/sQ4_K_M |
| Gemma 3 27B | Offload~5.5 tok/sQ4_K_M | Offload~3.6 tok/sQ4_K_M |
| Gemma 4 31B | Offload~4.6 tok/sQ4_K_M | Offload~3.1 tok/sQ4_K_M |
| Qwen3 32B | Offload~4.5 tok/sQ4_K_M | Offload~3.0 tok/sQ4_K_M |
| DeepSeek-R1 Distill 32B | Offload~4.5 tok/sQ4_K_M | Offload~3.0 tok/sQ4_K_M |
| Qwen2.5-Coder 32B | Offload~4.5 tok/sQ4_K_M | Offload~3.0 tok/sQ4_K_M |
| Qwen3.6 35B-A3B (MoE) | Offload~19 tok/sQ4_K_M | Offload~14 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 3090 / 3090 Ti is the better card for local AI. It runs 28 of 39 models fully in memory, against 21 for the RTX 4080 / 4080 Super, thanks to 24 GB of VRAM versus 16 GB. It is also about 14% faster on the models both can run.
Only the RTX 3090 / 3090 Ti 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 3090 / 3090 Ti and RTX 4080 / 4080 Super.
Questions
Is the RTX 3090 / 3090 Ti or the RTX 4080 / 4080 Super better for running LLMs?
The RTX 3090 / 3090 Ti is the better card for local AI. It runs 28 of 39 models fully in memory, against 21 for the RTX 4080 / 4080 Super, thanks to 24 GB of VRAM versus 16 GB. It is also about 14% faster on the models both can run.
Which models can the RTX 3090 / 3090 Ti run that the RTX 4080 / 4080 Super can't?
At 8K context, only the RTX 3090 / 3090 Ti 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 3090 / 3090 Ti or RTX 4080 / 4080 Super 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.