AMD · consumer

Radeon RX 6500 XT

Radeon RX 6500 XT has 4 GB of VRAM at 144 GB/s — about 3.72 GiB usable after driver and compositor overhead. 345 of 2118 indexed models fit at 64K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
4 GB
GDDR6
Bandwidth
144 GB/s
64-bit bus
Tensor FP16
—
dense
TDP
107 W
$199 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 253vision language 31audio tts 17video 2embedding 13audio asr 29

What fits at 64K context

largest quantization that fits, per model · 345 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Vikhr-Gemma-2B-instructIQ2_S2.6B0.96 GiB1.85 GiB3.72 GiB0.00 GiB32±26.5%
Gemmasutra-Mini-2B-v1I1-IQ2_S2.6B0.96 GiB1.85 GiB3.72 GiB0.00 GiB32±26.5%
Darwin-4B-ChimeraQ2_K_L4.0B1.73 GiB1.08 GiB3.72 GiB0.00 GiB32±26.5%
LocateAnything-3BQ3_K_M3.8B1.61 GiB1.20 GiB3.72 GiB0.00 GiB32±26.5%
EXAONE-4.0-1.2B-abliteratedI1-IQ4_NL1.5B0.84 GiB1.99 GiB3.71 GiB0.01 GiB32±26.5%
Trinity-Nano-PreviewMoEQ2_K_L6.1B2.32 GiB0.52 GiB3.71 GiB0.01 GiB61±37%
starcoder2-3bKV unresolvedQ4_13.0B1.80 GiB1.00 GiB3.71 GiB0.01 GiB32±26.5%
Qwen2.5-3BQ3_K_M3.1B1.61 GiB1.20 GiB3.71 GiB0.01 GiB32±26.5%
GRM-Kerlin-3bI1-Q3_K_M3.4B1.61 GiB1.20 GiB3.71 GiB0.01 GiB32±26.5%
Qwen2.5-Coder-3B-InstructQ3_K_M3.1B1.61 GiB1.20 GiB3.71 GiB0.01 GiB32±26.5%
Qwen2.5-3B-InstructQ3_K_M3.1B1.61 GiB1.20 GiB3.71 GiB0.01 GiB32±26.5%
granite-4.0-7B-A1B-Creative-v0.1MoEI1-IQ3_XS6.7B2.58 GiB0.27 GiB3.71 GiB0.01 GiB73±37%
Garnet-OCR-3B-0422I1-Q3_K_M4.1B1.61 GiB1.20 GiB3.71 GiB0.01 GiB32±26.5%
LFM2.5-Queen-Opus-4.7-8B-A1BMoEI1-IQ2_S8.5B2.41 GiB0.40 GiB3.71 GiB0.01 GiB59±37%
LFM2.5-8B-A1B-KO-SFTMoEI1-IQ2_S8.5B2.41 GiB0.40 GiB3.71 GiB0.01 GiB59±37%
LFM2.5-8B-A1B-hereticMoEI1-IQ2_S8.5B2.41 GiB0.40 GiB3.71 GiB0.01 GiB59±37%
LFM2.5-8B-A1B-SOMPOA-heresyMoEI1-IQ2_S8.5B2.41 GiB0.40 GiB3.71 GiB0.01 GiB59±37%
Huihui-LFM2.5-8B-A1B-abliteratedMoEI1-IQ2_S8.5B2.41 GiB0.40 GiB3.71 GiB0.01 GiB59±37%
Supertron2.1-8B-A1BMoEI1-IQ2_S8.5B2.41 GiB0.40 GiB3.71 GiB0.01 GiB59±37%
DeepSeek-OCR-2MoEQ4_K_M3.4B1.82 GiB1.00 GiB3.70 GiB0.02 GiB40±37%
DeepSeek-OCRMoEQ4_K_M3.3B1.82 GiB1.00 GiB3.70 GiB0.02 GiB40±37%
dolphin-2_6-phi-2Q8_02.8B2.75 GiB0.00 GiB3.70 GiB0.02 GiB32±26.5%
Qwen3.5-4B-NSFW-ARA-Heretic-LiteroticaI1-Q2_K_S4.2B1.73 GiB1.06 GiB3.70 GiB0.02 GiB32±26.5%
Qwen3.5-4B-RpRMax-v1I1-Q2_K_S4.7B1.73 GiB1.06 GiB3.70 GiB0.02 GiB32±26.5%
Holo-3.1-4B-uncensored-hereticI1-Q2_K_S4.5B1.73 GiB1.06 GiB3.70 GiB0.02 GiB32±26.5%
GRaPE-2-MiniI1-Q2_K_S4.7B1.73 GiB1.06 GiB3.70 GiB0.02 GiB32±26.5%
Qwen3.5-DPO-4B-2I1-Q2_K_S4.2B1.73 GiB1.06 GiB3.70 GiB0.02 GiB32±26.5%
Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliteratedI1-Q2_K_S4.7B1.73 GiB1.06 GiB3.70 GiB0.02 GiB32±26.5%
Qwopus3.5-4B-v3-hereticI1-Q2_K_S4.5B1.73 GiB1.06 GiB3.70 GiB0.02 GiB32±26.5%
Aureth-4B-Qwen3.5I1-Q2_K_S4.5B1.73 GiB1.06 GiB3.70 GiB0.02 GiB32±26.5%
Huihui-MiniCPM5-1B-abliteratedF161.1B2.02 GiB0.80 GiB3.70 GiB0.02 GiB32±26.5%
MiniCPM5-1B-Claude-Opus-Fable5-V2-ThinkingF161.1B2.02 GiB0.80 GiB3.70 GiB0.02 GiB32±26.5%
MiniCPM5-1B-UncensoredF161.1B2.02 GiB0.80 GiB3.70 GiB0.02 GiB32±26.5%
MiniCPM5-1B-Claude-Opus-Fable5-Thinking-abliteratedF161.1B2.02 GiB0.80 GiB3.70 GiB0.02 GiB32±26.5%
MiniCPM5-1B-Claude-Opus-Fable5-ThinkingF161.1B2.02 GiB0.80 GiB3.70 GiB0.02 GiB32±26.5%
MiniCPM5-1B-Agentic-Tooluse-Merged-FP16F161.1B2.02 GiB0.80 GiB3.70 GiB0.02 GiB32±26.5%
MiniCPM5-1BF161.1B2.02 GiB0.80 GiB3.70 GiB0.02 GiB32±26.5%
Dolphin3.0-Qwen2.5-3bQ3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
Qwen2.5-Coder-3B-Instruct-abliteratedI1-Q3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
GRM-Kerlin-3b-AbliteratedI1-Q3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
Mythos-nanoI1-Q3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
MATE-3BI1-Q3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
Mythos-nano-OBLITERATEDI1-Q3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
Qwen2.5-3B-Instruct-UncensoredI1-Q3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
Qwen2.5-Coder-3BQ3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
raspberry-3BQ3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
VibeThinker-3B-OBLITERATEDI1-Q3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
VibeThinker-3BQ3_K_L3.1B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
Fourier-Qwen2.5-VL-3B-0.67I1-Q3_K_L3.8B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
Qwen2.5-VL-3B-InstructQ3_K_L3.8B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
Nanonets-OCR-sI1-Q3_K_L3.8B1.59 GiB1.20 GiB3.70 GiB0.02 GiB32±26.5%
LFM2.5-8B-A1BMoEUD-IQ1_M8.5B2.40 GiB0.40 GiB3.69 GiB0.03 GiB59±37%
gemma-2-2b-itIQ2_XS2.6B0.93 GiB1.85 GiB3.69 GiB0.03 GiB32±26.5%
GLM-OCRI1-Q6_K1.3B0.68 GiB2.13 GiB3.69 GiB0.03 GiB32±26.5%
Gemma-3-4b-it-Uncensored-DBL-XI1-IQ3_XS4.7B1.88 GiB0.90 GiB3.69 GiB0.03 GiB32±26.5%
Holo-3.1-4BI1-IQ2_XS5.2B1.71 GiB1.06 GiB3.69 GiB0.03 GiB32±26.5%
AfriqueQwen3.5-4BI1-IQ2_XS5.2B1.71 GiB1.06 GiB3.69 GiB0.03 GiB32±26.5%
TimeOmni-1-4BI1-IQ2_XS5.2B1.71 GiB1.06 GiB3.69 GiB0.03 GiB32±26.5%
Qwen3.5-2B-hereticQ8_02.2B2.39 GiB0.40 GiB3.68 GiB0.04 GiB32±26.5%
Qwen2.5-3B-Instruct-abliteratedI1-IQ1_S3.1B1.57 GiB1.20 GiB3.68 GiB0.04 GiB32±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.

Questions people ask

What AI models can a Radeon RX 6500 XT run?
345 of 2118 indexed open-weight models fit a Radeon RX 6500 XT at 65,536 context with q8_0 KV cache, the largest being Vikhr-Gemma-2B-instruct at IQ2_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon RX 6500 XT actually have?
Its nameplate is 4 GB, but about 3.72 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon RX 6500 XT fast for local AI?
Its memory bandwidth is 144 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.