Compare · GPUs
RTX 4070 / 4070 Super vs RTX 3090 / 3090 Ti for local LLMs
It runs 28 of 39 models fully in memory, against 16 for the RTX 4070 / 4070 Super, thanks to 24 GB of VRAM versus 12 GB. It is also about 70% faster on the models both can run.
It runs 28 of 39 models fully in memory, against 16 for the RTX 4070 / 4070 Super, thanks to 24 GB of VRAM versus 12 GB. It is also about 70% 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 4070 / 4070 Super | RTX 3090 / 3090 Ti |
|---|---|---|
| Gemma 3 12B | Offload~6.8 tok/sQ4_K_M | Runs~47 tok/sQ8_0 |
| Phi-4 14B | Offload~5.7 tok/sQ4_K_M | Runs~39 tok/sQ8_0 |
| Qwen3 14B | Offload~5.6 tok/sQ4_K_M | Runs~39 tok/sQ8_0 |
| DeepSeek-R1 Distill 14B | Offload~5.6 tok/sQ4_K_M | Runs~39 tok/sQ8_0 |
| gpt-oss 20B | Offload~11 tok/sQ4_K_M | Runs~90 tok/sQ6_K |
| Codestral 22B | Offload~3.4 tok/sQ4_K_M | Runs~39 tok/sQ5_K_M |
| Mistral Small 3 24B | Offload~3.1 tok/sQ4_K_M | Runs~36 tok/sQ5_K_M |
| Devstral Small 24B | Offload~3.1 tok/sQ4_K_M | Runs~36 tok/sQ5_K_M |
| Gemma 4 26B-A4B (MoE) | Offload~10 tok/sQ4_K_M | Runs~110 tok/sQ4_K_M |
| Qwen3.6 27B | Offload~2.7 tok/sQ4_K_M | Runs~38 tok/sQ4_K_M |
| Qwen3 30B-A3B (MoE) | Offload~11 tok/sQ4_K_M | Runs~120 tok/sQ4_K_M |
| Qwen3-Coder 30B-A3B | Offload~11 tok/sQ4_K_M | Runs~120 tok/sQ4_K_M |
| Llama 3.3 70B | Won’t fit | Offload~1.9 tok/sQ3_K_M |
| DeepSeek-R1 Distill 70B | Won’t fit | Offload~1.9 tok/sQ3_K_M |
| Qwen3 0.6B | Runs~520 tok/sQ8_0 | Runs~960 tok/sQ8_0 |
| Gemma 3 1B | Runs~310 tok/sQ8_0 | Runs~570 tok/sQ8_0 |
| Llama 3.2 1B | Runs~260 tok/sQ8_0 | Runs~480 tok/sQ8_0 |
| Llama 3.2 3B | Runs~97 tok/sQ8_0 | Runs~180 tok/sQ8_0 |
| Phi-4 Mini 3.8B | Runs~81 tok/sQ8_0 | Runs~150 tok/sQ8_0 |
| Qwen3 4B | Runs~77 tok/sQ8_0 | Runs~140 tok/sQ8_0 |
| Gemma 3 4B | Runs~72 tok/sQ8_0 | Runs~130 tok/sQ8_0 |
| Qwen3.5 4B | Runs~66 tok/sQ8_0 | Runs~120 tok/sQ8_0 |
| Mistral 7B v0.3 | Runs~43 tok/sQ8_0 | Runs~80 tok/sQ8_0 |
| Qwen2.5-Coder 7B | Runs~41 tok/sQ8_0 | Runs~76 tok/sQ8_0 |
| Llama 3.1 8B | Runs~39 tok/sQ8_0 | Runs~72 tok/sQ8_0 |
| DeepSeek-R1 Distill 8B | Runs~39 tok/sQ8_0 | Runs~72 tok/sQ8_0 |
| Qwen3 8B | Runs~38 tok/sQ8_0 | Runs~70 tok/sQ8_0 |
| Qwen3.5 9B | Runs~43 tok/sQ6_K | Runs~61 tok/sQ8_0 |
| Gemma 4 12B | Runs~46 tok/sQ4_K_M | Runs~48 tok/sQ8_0 |
| Mistral Nemo 12B | Runs~46 tok/sQ4_K_M | Runs~47 tok/sQ8_0 |
| Gemma 3 27B | Offload~2.6 tok/sQ4_K_M | Offload~5.5 tok/sQ4_K_M |
| Gemma 4 31B | Offload~2.3 tok/sQ4_K_M | Offload~4.6 tok/sQ4_K_M |
| Qwen3 32B | Offload~2.1 tok/sQ4_K_M | Offload~4.5 tok/sQ4_K_M |
| DeepSeek-R1 Distill 32B | Offload~2.1 tok/sQ4_K_M | Offload~4.5 tok/sQ4_K_M |
| Qwen2.5-Coder 32B | Offload~2.1 tok/sQ4_K_M | Offload~4.5 tok/sQ4_K_M |
| Qwen3.6 35B-A3B (MoE) | Offload~12 tok/sQ4_K_M | Offload~19 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 16 for the RTX 4070 / 4070 Super, thanks to 24 GB of VRAM versus 12 GB. It is also about 70% faster on the models both can run.
Only the RTX 3090 / 3090 Ti fits Gemma 3 12B, Phi-4 14B, Qwen3 14B, DeepSeek-R1 Distill 14B, gpt-oss 20B, Codestral 22B, Mistral Small 3 24B and Devstral Small 24B and 4 more fully in memory.
See every result on its own page: RTX 4070 / 4070 Super and RTX 3090 / 3090 Ti.
Questions
Is the RTX 4070 / 4070 Super or the RTX 3090 / 3090 Ti 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 16 for the RTX 4070 / 4070 Super, thanks to 24 GB of VRAM versus 12 GB. It is also about 70% faster on the models both can run.
Which models can the RTX 3090 / 3090 Ti run that the RTX 4070 / 4070 Super can't?
At 8K context, only the RTX 3090 / 3090 Ti holds Gemma 3 12B, Phi-4 14B, Qwen3 14B, DeepSeek-R1 Distill 14B, gpt-oss 20B and Codestral 22B and 6 more fully in VRAM.
Does the RTX 4070 / 4070 Super or RTX 3090 / 3090 Ti 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.