AMD · consumer

Radeon RX 7900 XT

Radeon RX 7900 XT has 20 GB of VRAM at 800 GB/s — about 18.60 GiB usable after driver and compositor overhead. 1947 of 2118 indexed models fit at 16K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
20 GB
GDDR6
Bandwidth
800 GB/s
320-bit bus
Tensor FP16
—
dense
TDP
315 W
$899 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1670vision language 173video 16audio asr 39image 2audio tts 21embedding 26

What fits at 16K context

largest quantization that fits, per model · 1947 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
cogito-v1-preview-qwen-32BI1-IQ4_XS32.8B16.48 GiB1.13 GiB18.60 GiB0.00 GiB28±26.5%
QwQ-32B-Snowdrop-v0I1-IQ4_XS32.8B16.48 GiB1.13 GiB18.60 GiB0.00 GiB28±26.5%
DeepSeek-R1-Distill-Qwen-32B-UncensoredI1-IQ4_XS32.8B16.48 GiB1.13 GiB18.60 GiB0.00 GiB28±26.5%
RoguePlanet-DeepSeek-R1-Qwen-32B-RPI1-IQ4_XS32.8B16.48 GiB1.13 GiB18.60 GiB0.00 GiB28±26.5%
Olmo-3.1-32B-InstructQ4_032.2B17.08 GiB0.52 GiB18.60 GiB0.00 GiB28±26.5%
Olmo-3.1-32B-ThinkQ4_032.2B17.08 GiB0.52 GiB18.60 GiB0.00 GiB28±26.5%
Olmo-3-32B-ThinkQ4_032.2B17.08 GiB0.52 GiB18.60 GiB0.00 GiB28±26.5%
Salience-1.5-FlashMoEI1-Q4_K_M31.1B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Huihui-Qwen3-VL-30B-A3B-Instruct-abliteratedMoEI1-Q4_K_M31.1B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Qwen3-VL-30B-A3B-InstructMoEQ4_K_M31.1B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Qwen3-30B-A3B-Gemini-Pro-High-Reasoning-2507-ABLITERATED-UNCENSOREDMoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
MiroThinker-v1.0-30BMoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Qwen3-30B-A3B-YOYO-V5MoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Qwen3-30B-A3B-Thinking-2507-Claude-4.5-Sonnet-High-Reasoning-DistillMoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Huihui-Qwen3-30B-A3B-Thinking-2507-abliteratedMoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Huihui-Qwen3-30B-A3B-Instruct-2507-abliteratedMoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Qwen3-30B-A3B-abliterated-eroticMoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Qwen3-30B-A3B-abliteratedMoEQ4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Huihui-Qwen3-Coder-30B-A3B-Instruct-abliteratedMoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Qwen3-Coder-30B-A3B-Instruct-RTPurboMoEI1-Q4_K_M30.5B17.28 GiB0.42 GiB18.60 GiB0.00 GiB102±37%
Qwen3-VL-32B-Instruct-ultra-uncensored-hereticI1-IQ4_XS33.4B16.48 GiB1.13 GiB18.59 GiB0.01 GiB28±26.5%
Huihui-Qwen3-VL-32B-Instruct-abliteratedI1-IQ4_XS33.4B16.48 GiB1.13 GiB18.59 GiB0.01 GiB28±26.5%
ColorGUI-32BI1-IQ4_XS33.4B16.48 GiB1.13 GiB18.59 GiB0.01 GiB28±26.5%
Qwen3-32B-UncensoredI1-IQ4_XS32.8B16.48 GiB1.13 GiB18.59 GiB0.01 GiB28±26.5%
Qwen3-32B-abliteratedI1-IQ4_XS32.8B16.48 GiB1.13 GiB18.59 GiB0.01 GiB28±26.5%
DeepSWE-PreviewIQ4_XS32.8B16.48 GiB1.13 GiB18.59 GiB0.01 GiB28±26.5%
AReaL-boba-2-32BI1-IQ4_XS32.8B16.48 GiB1.13 GiB18.59 GiB0.01 GiB28±26.5%
Assistant_Pepe_32BI1-IQ4_XS32.8B16.48 GiB1.13 GiB18.59 GiB0.01 GiB28±26.5%
Trinity-MiniMoEQ5_K_L26.1B17.59 GiB0.10 GiB18.59 GiB0.01 GiB113±37%
IQuest-Coder-V1-40B-InstructI1-IQ3_S39.8B16.17 GiB1.41 GiB18.58 GiB0.02 GiB28±26.5%
Phi-3.5-MoE-instructMoEKV unresolvedIQ3_M41.9B17.11 GiB0.56 GiB18.57 GiB0.03 GiB73±37%
GLM-Z1-32B-0414IQ4_NL32.6B17.31 GiB0.27 GiB18.57 GiB0.03 GiB28±26.5%
GLM-4-32B-0414IQ4_NL32.6B17.31 GiB0.27 GiB18.57 GiB0.03 GiB28±26.5%
Huihui-Qwen3.5-35B-A3B-abliteratedMoEIQ4_XS36.0B17.57 GiB0.09 GiB18.56 GiB0.04 GiB141±37%
Qwen3.5-35B-A3B-BaseMoEIQ4_XS36.0B17.57 GiB0.09 GiB18.56 GiB0.04 GiB141±37%
Gemma-4-Gembrain-X-Core-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma-4-Gembrain-X-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma-4-31B-Isometry-Fabled-PersonaI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Versipellis-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma4-Gutenberg-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
G4-MeroMero-31B-uncensored-hereticI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma-4-Novelist-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Wanabi-Gemma4-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
G4-Alice-v1.2-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Agares-31B-v1I1-Q4_K_S30.7B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma4-Gutenberg-31B-HereticI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-hereticI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma-4-Gemsicle-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma-4-Gembrain-31B-it-uncensored-hereticI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Melinoe-Gemma4-31B-VL-hereticI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
G4-MeroMero-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Glistening-Gem-31B-v1.0I1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Melinoe-Gemma4-31B-VLI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma-4-31B-Storymaxxed3I1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Huihui-gemma-4-31B-it-qat-q4_0-unquantized-abliteratedI1-Q4_K_S32.7B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
gemma-4-31B-Queen-it-qat-q4_0-unquantizedI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
gemma-4-31B-it-qat-q4_0-unquantized-hereticI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
Gemma-4-AssGuard-31BI1-Q4_K_S31.3B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
copywriter-gemma4-31bI1-Q4_K_S32.7B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
gemma-4-31B-heretic-finetuneI1-Q4_K_S30.7B16.54 GiB1.03 GiB18.56 GiB0.04 GiB28±26.5%
●From the file◐Predictedwhat 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
Workload◍MedianMiddle 50%Runs
Image generation11.45 it/s7.75–16.22328
Prompt processing3219.16 tok/s2738.95–3754.6863
Text generation101.20 tok/s99.80–107.4539
Benchmarked· n=328

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 7900 XT run?
1947 of 2118 indexed open-weight models fit a Radeon RX 7900 XT at 16,384 context with q4_0 KV cache, the largest being cogito-v1-preview-qwen-32B at I1-IQ4_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon RX 7900 XT actually have?
Its nameplate is 20 GB, but about 18.60 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon RX 7900 XT fast for local AI?
Its memory bandwidth is 800 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.