Skip to content

AMD Radeon RX 9070

AMD · 16GBGDDR6 · Can run 81 models

BuyAmazon
ManufacturerAMD
VRAM16 GB
Memory TypeGDDR6
ArchitectureRDNA 4
Stream Procs3,584
Bandwidth538 GB/s
TDP250W
MSRP$449
ReleasedJan 23, 2025

AI Notes

The RX 9070 brings RDNA 4 to a more accessible price with 16GB VRAM. It can run 13B models and attempt larger ones with quantization. The lower bandwidth compared to the 9070 XT means slightly slower token generation, but the VRAM capacity is identical. A compelling option if ROCm support matures for RDNA 4.

Compatible Models

ModelParametersBest QuantVRAM UsedFitEst. Speed
Qwen 3 0.6B600MQ4_K_M2.5 GBRuns~215 tok/s
Qwen 3.5 0.8B800MQ4_K_M1.5 GBRuns~359 tok/s
Gemma 3 1B1BQ8_02 GBRuns~269 tok/s
Llama 3.2 1B1BQ8_03 GBRuns~179 tok/s
DeepSeek R1 1.5B1.5BQ8_03 GBRuns~179 tok/s
SmolLM2 1.7B1.7BQ8_02.7 GBRuns~199 tok/s
Gemma 2 2B2BQ8_04 GBRuns~135 tok/s
Gemma 3n E2B2BQ4_K_M3.3 GBRuns~163 tok/s
Gemma 4 E2B2BQ4_K_M4 GBRuns~135 tok/s
Qwen 3.5 2B2BQ4_K_M3 GBRuns~179 tok/s
Llama 3.2 3B3BQ8_05 GBRuns~108 tok/s
StarCoder2 3B3BQ4_K_M3.5 GBRuns~154 tok/s
Phi-3 Mini 3.8B3.8BQ8_05.8 GBRuns~93 tok/s
Phi-4 Mini 3.8B3.8BQ4_K_M4.5 GBRuns~120 tok/s
Gemma 3 4B4BQ4_K_M5 GBRuns~108 tok/s
Gemma 3n E4B4BQ4_K_M4.5 GBRuns~120 tok/s
Gemma 4 E4B4BQ4_K_M6 GBRuns~90 tok/s
Qwen 3 4B4BQ4_K_M4.5 GBRuns~120 tok/s
Qwen 3.5 4B4BQ4_K_M4.5 GBRuns~120 tok/s
Yi 1.5 6B6BQ4_K_M5 GBRuns~108 tok/s
Codestral Mamba 7B7BQ4_K_M6.9 GBRuns~78 tok/s
DeepSeek R1 7B7BQ8_09 GBRuns~60 tok/s
Falcon 3 7B7BQ4_K_M6.8 GBRuns~79 tok/s
InternLM 2.5 7B7BQ4_K_M5.5 GBRuns~98 tok/s
Mistral 7B7BQ8_09 GBRuns~60 tok/s
OpenChat 3.5 7B7BQ4_K_M6.9 GBRuns~78 tok/s
Qwen 2.5 7B7BQ8_09 GBRuns~60 tok/s
Qwen 2.5 Coder 7B7BQ8_09 GBRuns~60 tok/s
Qwen 2.5 VL 7B7BQ4_K_M7 GBRuns~77 tok/s
StarCoder2 7B7BQ4_K_M5.5 GBRuns~98 tok/s
WizardLM 2 7B7BQ4_K_M6.9 GBRuns~78 tok/s
Aya Expanse 8B8BQ4_K_M6.5 GBRuns~83 tok/s
Cogito 8B8BQ4_K_M7.5 GBRuns~72 tok/s
DeepSeek R1 8B8BQ4_K_M7.5 GBRuns~72 tok/s
Dolphin 3 8B8BQ4_K_M6 GBRuns~90 tok/s
Granite 3.3 8B8BQ8_010 GBRuns~54 tok/s
Llama 3.1 8B8BQ8_010 GBRuns~54 tok/s
Nemotron 3 Nano 8B8BQ4_K_M7.5 GBRuns~72 tok/s
Nous Hermes 2 8B8BQ4_K_M6 GBRuns~90 tok/s
Qwen 3 8B8BQ4_K_M7.5 GBRuns~72 tok/s
Gemma 2 9B9BQ8_011 GBRuns~49 tok/s
Qwen 3.5 9B9BQ4_K_M7.5 GBRuns~72 tok/s
Yi 1.5 9B9BQ4_K_M6.5 GBRuns~83 tok/s
Yi Coder 9B9BQ4_K_M8 GBRuns~67 tok/s
Falcon 3 10B10BQ4_K_M8.5 GBRuns~63 tok/s
Llama 3.2 Vision 11B11BQ4_K_M8.5 GBRuns~63 tok/s
Gemma 3 12B12BQ4_K_M10.5 GBRuns~51 tok/s
Mistral Nemo 12B12BQ4_K_M9.5 GBRuns~57 tok/s
DeepSeek R1 14B14BQ4_K_M9.9 GBRuns~54 tok/s
Phi-4 14B14BQ4_K_M9.9 GBRuns~54 tok/s
Phi-4 Reasoning 14B14BQ4_K_M11 GBRuns~49 tok/s
Qwen 2.5 14B14BQ4_K_M9.9 GBRuns~54 tok/s
Qwen 2.5 Coder 14B14BQ4_K_M12 GBRuns~45 tok/s
Qwen 3 14B14BQ4_K_M12 GBRuns~45 tok/s
InternLM 2.5 20B20BQ4_K_M12 GBRuns~45 tok/s
Qwen 3.5 35B A3B35BQ4_K_M12 GBRuns~45 tok/s
gpt-oss 20B21BMXFP415 GBRuns (tight)~36 tok/s
Codestral 22B22BQ4_K_M14.7 GBRuns (tight)~37 tok/s
StarCoder2 15B15BQ8_017 GBCPU Offload~10 tok/s
Devstral 24B24BQ4_K_M17 GBCPU Offload~10 tok/s
Magistral Small 24B24BQ4_K_M17 GBCPU Offload~10 tok/s
Mistral Small 3.1 24B24BQ4_K_M18 GBCPU Offload~9 tok/s
Gemma 4 26B26BQ4_K_M20 GBCPU Offload~8 tok/s
Gemma 2 27B27BQ4_K_M17.7 GBCPU Offload~9 tok/s
Gemma 3 27B27BQ4_K_M20 GBCPU Offload~8 tok/s
Qwen 3.5 27B27BQ4_K_M19 GBCPU Offload~8 tok/s
Qwen 3.6 27B27BQ4_K_M20 GBCPU Offload~8 tok/s
Qwen 3 30B-A3B (MoE)30BQ4_K_M22 GBCPU Offload~7 tok/s
Gemma 4 31B31BQ4_K_M22 GBCPU Offload~7 tok/s
Aya Expanse 32B32BQ4_K_M22 GBCPU Offload~7 tok/s
Cogito 32B32BQ4_K_M21.5 GBCPU Offload~8 tok/s
DeepSeek R1 32B32BQ4_K_M20.7 GBCPU Offload~8 tok/s
Qwen 2.5 32B32BQ4_K_M20.7 GBCPU Offload~8 tok/s
Qwen 2.5 Coder 32B32BQ4_K_M23 GBCPU Offload~7 tok/s
Qwen 3 32B32BQ4_K_M23 GBCPU Offload~7 tok/s
QwQ 32B32BQ4_K_M21.5 GBCPU Offload~8 tok/s
Laguna XS 2.133BQ4_K_M22 GBCPU Offload~7 tok/s
WizardCoder 33B33BQ4_K_M22 GBCPU Offload~7 tok/s
Nous Hermes 2 34B34BQ4_K_M19 GBCPU Offload~8 tok/s
Yi 1.5 34B34BQ4_K_M21 GBCPU Offload~8 tok/s
Command R 35B35BQ4_K_M22.5 GBCPU Offload~7 tok/s
33 model(s) are too large for this hardware.