Radeon RX 7900 XTX vs RTX 4090 for local LLMs

Both run the same 48 of 69 models well. On models that fit both, the RTX 4090 should be about 5% faster, because generation speed follows memory bandwidth (1008 vs 960 GB/s).

AMD Radeon RX 7900 XTX

  • 24 GB VRAM
  • 960 GB/s bandwidth
  • 48/69 models run well
  • $999 at launch

NVIDIA RTX 4090

  • 24 GB VRAM
  • 1008 GB/s bandwidth
  • 48/69 models run well
  • $1,599 at launch

Prices are US launch prices (MSRP) for comparison; street prices change daily. Links go to Amazon, where we earn a commission at no cost to you.

Every model, side by side

Best quant that fits and estimated speed at 8K context.

ModelRadeon RX 7900 XTXRTX 4090
Qwen3 0.6BFP16 · ~384 tok/sFP16 · ~403 tok/s
Gemma 3 1BFP16 · ~288 tok/sFP16 · ~302 tok/s
Llama 3.2 1BFP16 · ~232 tok/sFP16 · ~244 tok/s
Qwen3 1.7BFP16 · ~142 tok/sFP16 · ~149 tok/s
Llama 3.2 3BFP16 · ~90 tok/sFP16 · ~94 tok/s
Phi-3.5 Mini 3.8BFP16 · ~75 tok/sFP16 · ~79 tok/s
Phi-4 Mini 3.8BFP16 · ~75 tok/sFP16 · ~79 tok/s
Qwen3 4BFP16 · ~72 tok/sFP16 · ~75 tok/s
Gemma 3 4BFP16 · ~67 tok/sFP16 · ~70 tok/s
Mistral 7B v0.3FP16 · ~40 tok/sFP16 · ~42 tok/s
Qwen2.5 7BFP16 · ~38 tok/sFP16 · ~40 tok/s
Qwen2.5 Coder 7BFP16 · ~38 tok/sFP16 · ~40 tok/s
DeepSeek R1 Distill Qwen 7BFP16 · ~38 tok/sFP16 · ~40 tok/s
Llama 3.1 8BFP16 · ~36 tok/sFP16 · ~38 tok/s
Granite 3.3 8BFP16 · ~35 tok/sFP16 · ~37 tok/s
Qwen3 8BFP16 · ~35 tok/sFP16 · ~37 tok/s
DeepSeek R1 0528 Qwen3 8BFP16 · ~35 tok/sFP16 · ~37 tok/s
Gemma 2 9BFP16 · ~31 tok/sFP16 · ~33 tok/s
Gemma 4 12BQ8_0 · ~45 tok/sQ8_0 · ~48 tok/s
Mistral Nemo 12BQ8_0 · ~44 tok/sQ8_0 · ~47 tok/s
Gemma 3 12BQ8_0 · ~44 tok/sQ8_0 · ~47 tok/s
Qwen2.5 14BQ8_0 · ~37 tok/sQ8_0 · ~39 tok/s
Phi-4 14BQ8_0 · ~37 tok/sQ8_0 · ~39 tok/s
Phi-4 Reasoning 14BQ8_0 · ~37 tok/sQ8_0 · ~39 tok/s
Qwen3 14BQ8_0 · ~37 tok/sQ8_0 · ~38 tok/s
Qwen2.5 Coder 14BQ8_0 · ~37 tok/sQ8_0 · ~38 tok/s
DeepSeek R1 Distill Qwen 14BQ8_0 · ~37 tok/sQ8_0 · ~38 tok/s
gpt-oss 20BQ8_0 · ~151 tok/sQ8_0 · ~158 tok/s
Mistral Small 24BQ6_K · ~30 tok/sQ6_K · ~31 tok/s
Devstral Small 24BQ6_K · ~30 tok/sQ6_K · ~31 tok/s
Magistral Small 24BQ6_K · ~30 tok/sQ6_K · ~31 tok/s
Mistral Small 3.2 24BQ6_K · ~29 tok/sQ6_K · ~31 tok/s
Gemma 4 26B-A4BQ6_K · ~185 tok/sQ6_K · ~194 tok/s
Gemma 2 27BQ5_K_M · ~30 tok/sQ5_K_M · ~31 tok/s
Gemma 3 27BQ5_K_M · ~30 tok/sQ5_K_M · ~31 tok/s
Qwen3.8 27BQ6_K · ~25 tok/sQ6_K · ~27 tok/s
Qwen3.6 27BQ6_K · ~25 tok/sQ6_K · ~27 tok/s
Qwen3 30B-A3BQ5_K_M · ~245 tok/sQ5_K_M · ~258 tok/s
Qwen3 Coder 30B-A3BQ5_K_M · ~245 tok/sQ5_K_M · ~258 tok/s
GLM-4.7 Flash 30B-A3BQ5_K_M · ~270 tok/sQ5_K_M · ~283 tok/s
Gemma 4 31BQ4_K_M · ~30 tok/sQ4_K_M · ~32 tok/s
Nemotron 3 Nano 30B-A3BQ5_K_M · ~231 tok/sQ5_K_M · ~243 tok/s
Qwen2.5 32BQ4_K_M · ~29 tok/sQ4_K_M · ~30 tok/s
DeepSeek R1 Distill Qwen 32BQ4_K_M · ~29 tok/sQ4_K_M · ~30 tok/s
Qwen3 32BQ4_K_M · ~29 tok/sQ4_K_M · ~30 tok/s
QwQ 32BQ4_K_M · ~29 tok/sQ4_K_M · ~30 tok/s
Qwen2.5 Coder 32BQ4_K_M · ~29 tok/sQ4_K_M · ~30 tok/s
Qwen3.6 35B-A3BQ4_K_M · ~317 tok/sQ4_K_M · ~333 tok/s
Mixtral 8x7BQ2_K · ~119 tok/sQ2_K · ~125 tok/s
Llama 3.1 70BToo bigToo big
Llama 3.3 70BToo bigToo big
DeepSeek R1 Distill Llama 70BToo bigToo big
Qwen2.5 72BToo bigToo big
Qwen3 Coder Next 80B-A3BToo bigToo big
GLM-4.5-Air 106B-A12BToo bigToo big
Llama 4 Scout 17B-16EToo bigToo big
gpt-oss 120BToo bigToo big
Qwen3.8 Flash NextToo bigToo big
Qwen3 235B-A22BToo bigToo big
DeepSeek V4 Flash 284B-A13BToo bigToo big
GLM-5.3 Flash 320B-A18BToo bigToo big
MiniMax M3 428B-A23BToo bigToo big
GLM-5 744B-A40BToo bigToo big
GLM-5.3Too bigToo big
DeepSeek V4.1 FlashToo bigToo big
Mistral Large 4 1T-A52BToo bigToo big
DeepSeek V4 Pro 1.6T-A49BToo bigToo big
Qwen3.8 2.4T-A95BToo bigToo big
Kimi K3.1 Pro 2.8T-A104BToo bigToo big