AMD · consumer

Radeon RX 9070 XT

Radeon RX 9070 XT has 16 GB of VRAM at 640 GB/s — about 14.88 GiB usable after driver and compositor overhead. 1811 of 2118 indexed models fit at 64K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
16 GB
GDDR6
Bandwidth
640 GB/s
256-bit bus
Tensor FP16
dense
TDP
304 W
$599 MSRP
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1549vision language 159image 2video 15audio asr 39embedding 26audio tts 21

What fits at 64K context

largest quantization that fits, per model · 1811 of 2118 indexed
ModelBest quantParamsWeightsKVTotal in memoryHeadroomtok/s
Llama3.2-24B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-Q5_K_M18.0B12.00 GiB1.97 GiB14.88 GiB0.00 GiB53±37%
GRM-2.6-Plus-0628Q3_K_S27.8B12.78 GiB1.13 GiB14.87 GiB0.01 GiB29±26.5%
ThinkingCap-Qwen3.6-27BQ3_K_S27.4B12.78 GiB1.13 GiB14.87 GiB0.01 GiB29±26.5%
Tess-4-27BQ3_K_S27.8B12.78 GiB1.13 GiB14.87 GiB0.01 GiB29±26.5%
Qwen3-16B-A3BMoEQ6_K16.0B12.28 GiB1.69 GiB14.86 GiB0.02 GiB56±37%
Llama-3.2-3BF163.2B11.98 GiB1.97 GiB14.86 GiB0.02 GiB29±26.5%
Gemma-4-31B-Isometry-RPI1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Gemma-4-Dark-Gemistry-31BI1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Prosopon-31BI1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Gemma-4-Novelist-Eclipse-31BI1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Giftige-Blume-31B-v1-StyleSwapI1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
G4-MeroMero-31B-StyleSwapI1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Gemma-4-31B-StyleTune-heretic-araI1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Pantheon-Reasoning-31B-1.1I1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Gemma-4-31B-StyleTuneI1-IQ2_M32.7B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Barcenas-StyleTune-31B-FableI1-IQ2_M32.1B10.73 GiB3.14 GiB14.85 GiB0.03 GiB29±26.5%
Dolphin3.0-Llama3.2-3BF323.2B11.98 GiB1.97 GiB14.85 GiB0.03 GiB29±26.5%
Hermes-3-Llama-3.2-3BF323.2B11.98 GiB1.97 GiB14.85 GiB0.03 GiB29±26.5%
Tiger-Gemma-12B-v3Q8_012.8B12.65 GiB1.26 GiB14.85 GiB0.03 GiB29±26.5%
Goetia-26B-A4B-v1.4MoEI1-Q3_K_L26.0B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
G4-Moonlight-Dusk-26B-A4B-hereticMoEI1-Q3_K_L26.5B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
Pantheon-Reasoning-26B-A4B-1.1-hereticMoEI1-Q3_K_L26.5B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
G4-Moonlight-Dusk-26B-A4BMoEI1-Q3_K_L26.5B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
Chimera-X-26B-A4BMoEI1-Q3_K_L26.5B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
Pantheon-Reasoning-26B-A4B-1.1MoEI1-Q3_K_L26.5B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
Gemma-4-26B-A4B-StyleTune-V2MoEI1-Q3_K_L26.5B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
Gemma-4-26B-A4B-StyleTuneMoEI1-Q3_K_L26.5B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
gemma-4-26b-a4b-heretic-styletune-v2-headMoEI1-Q3_K_L25.8B13.17 GiB0.79 GiB14.84 GiB0.04 GiB29±26.5%
Qwen3.6-VL-REAP-26B-A3BMoEIQ4_XS26.6B13.59 GiB0.35 GiB14.84 GiB0.04 GiB116±37%
granite-20b-code-instruct-8kQ5_K_L20.1B13.86 GiB0.00 GiB14.84 GiB0.04 GiB29±26.5%
Gemma4-Gutenberg-31BIQ2_XS31.3B10.71 GiB3.14 GiB14.83 GiB0.05 GiB29±26.5%
gemma-4-31B-itIQ2_XS31.3B10.71 GiB3.14 GiB14.83 GiB0.05 GiB29±26.5%
Gemma4-Gutenberg-31B-HereticIQ2_XS31.3B10.71 GiB3.14 GiB14.83 GiB0.05 GiB29±26.5%
Equinox-31BIQ2_XS31.3B10.71 GiB3.14 GiB14.83 GiB0.05 GiB29±26.5%
gemma-4-31B-it-SDFT-Heretic-RPIQ2_XS30.7B10.71 GiB3.14 GiB14.83 GiB0.05 GiB29±26.5%
ERNIE-4.5-21B-A3B-ThinkingQ4_121.8B12.93 GiB0.98 GiB14.83 GiB0.05 GiB29±26.5%
ERNIE-4.5-21B-A3B-PTQ4_121.9B12.93 GiB0.98 GiB14.83 GiB0.05 GiB29±26.5%
Phi-3.5-MoE-instructMoEKV unresolvedIQ2_S41.9B11.67 GiB2.25 GiB14.83 GiB0.05 GiB51±37%
spoomplesmaxx-v2.1-30BI1-Q2_K_S28.9B9.31 GiB4.50 GiB14.82 GiB0.06 GiB29±26.5%
Huihui-granite-4.1-30b-abliteratedI1-Q2_K_S28.9B9.31 GiB4.50 GiB14.82 GiB0.06 GiB29±26.5%
granite-4.1-30b-hereticI1-Q2_K_S28.9B9.31 GiB4.50 GiB14.82 GiB0.06 GiB29±26.5%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEI1-Q4_K_M23.0B12.98 GiB0.93 GiB14.82 GiB0.06 GiB77±37%
grug-27bIQ3_M27.4B12.73 GiB1.13 GiB14.81 GiB0.07 GiB29±26.5%
Carnice-V2-27bIQ3_M27.4B12.73 GiB1.13 GiB14.81 GiB0.07 GiB29±26.5%
Fara1.5-27BIQ3_M27.4B12.73 GiB1.13 GiB14.81 GiB0.07 GiB29±26.5%
Olmo-3.1-32B-InstructIQ3_XS32.2B12.45 GiB1.36 GiB14.81 GiB0.07 GiB29±26.5%
Olmo-3.1-32B-ThinkIQ3_XS32.2B12.45 GiB1.36 GiB14.81 GiB0.07 GiB29±26.5%
Olmo-3-32B-ThinkIQ3_XS32.2B12.45 GiB1.36 GiB14.81 GiB0.07 GiB29±26.5%
Llama3.2-30B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-UncensoredMoEI1-Q2_K30.0B10.60 GiB3.30 GiB14.81 GiB0.07 GiB43±37%
Yi-34B-200K-DARE-megamerge-v8I1-IQ2_XS34.4B9.60 GiB4.22 GiB14.80 GiB0.08 GiB29±26.5%
dolphin-2.9.1-yi-1.5-34bI1-IQ2_XS34.4B9.60 GiB4.22 GiB14.80 GiB0.08 GiB29±26.5%
OrionStar-Yi-34B-Chat-LlamaI1-IQ2_XS34.4B9.60 GiB4.22 GiB14.80 GiB0.08 GiB29±26.5%
Yi-34B-200K-LlamafiedI1-IQ2_XS34.4B9.60 GiB4.22 GiB14.80 GiB0.08 GiB29±26.5%
Yi-1.5-34BIQ2_XS34.4B9.60 GiB4.22 GiB14.80 GiB0.08 GiB29±26.5%
Merged-RP-Stew-V2-34BI1-IQ2_XS34.4B9.60 GiB4.22 GiB14.80 GiB0.08 GiB29±26.5%
Capybara-Tess-Yi-34B-200KI1-IQ2_XS34.4B9.60 GiB4.22 GiB14.80 GiB0.08 GiB29±26.5%
Gemma-3-27B-MeditronFOI1-IQ3_M28.8B12.25 GiB1.58 GiB14.80 GiB0.08 GiB29±26.5%
GLM-4-32B-0414-Korean-CultureI1-IQ3_XS32.6B12.72 GiB1.07 GiB14.78 GiB0.10 GiB29±26.5%
GLM-Z1-32B-0414IQ3_XS32.6B12.72 GiB1.07 GiB14.78 GiB0.10 GiB29±26.5%
GLM-4-32B-0414IQ3_XS32.6B12.72 GiB1.07 GiB14.78 GiB0.10 GiB29±26.5%
From the filePredictedwhat these mean

Speed is modeled, not measured: decode is memory-bandwidth bound, so tokens per second is bytes read per token against achievable bandwidth. Mixture-of-experts models carry a wider band because only the routed experts are read each step, and few have been measured publicly.

Measured on this card

third-party benchmarks, aggregated
WorkloadMedianMiddle 50%Runs
Image generation5.76 it/s2.5411.88118
Prompt processing4275.83 tok/s3591.414809.8730
Text generation95.24 tok/s86.48108.0130
Benchmarked· n=118

Aggregated from community-submitted runs, so the spread is wide by nature — it covers different models, resolutions, step counts and settings, not one controlled configuration. Read the middle 50% rather than the median alone. These figures are reproduced with attribution from vladmandic-sd-data-benchmark, which publishes no licence — so we display and link rather than redistribute them.

Questions people ask

What AI models can a Radeon RX 9070 XT run?
1811 of 2118 indexed open-weight models fit a Radeon RX 9070 XT at 65,536 context with q4_0 KV cache, the largest being Llama3.2-24B-A3B-II-Dark-Champion-INSTRUCT-Heretic-Abliterated-Uncensored at I1-Q5_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon RX 9070 XT actually have?
Its nameplate is 16 GB, but about 14.88 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon RX 9070 XT fast for local AI?
Its memory bandwidth is 640 GB/s, and that figure — not teraflops — is what governs token generation speed. Capacity decides what you can run; bandwidth decides how fast it runs.