AMD · consumer

Radeon RX 6650 XT

Radeon RX 6650 XT has 8 GB of VRAM at 280 GB/s — about 7.44 GiB usable after driver and compositor overhead. 431 of 2118 indexed models fit at 128K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
8 GB
GDDR6
Bandwidth
280 GB/s
128-bit bus
Tensor FP16
—
dense
TDP
180 W
$399 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 319vision language 43embedding 14audio tts 18video 7image 1audio asr 29

What fits at 128K context

largest quantization that fits, per model · 431 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Agents-A1-4B-Heretic-ARA-Refusals8Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4B-NSFW-ARA-Heretic-LiteroticaI1-Q4_K_M4.2B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Agents-A1-4BQ4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4B-RpRMax-v1I1-Q4_K_M4.7B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Holo-3.1-4B-uncensored-hereticI1-Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
GRaPE-2-MiniI1-Q4_K_M4.7B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-DPO-4B-2I1-Q4_K_M4.2B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4B-BaseQ4_K_M4.7B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4B-hereticQ4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
NuExtract3Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliteratedI1-Q4_K_M4.7B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwopus3.5-4B-v3-hereticI1-Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4BQ4_K_M4.7B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Aureth-4B-Qwen3.5I1-Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKINGI1-Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4B-SOMPOA-heresy-v2I1-Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4B-SOMPOA-heresyI1-Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Qwen3.5-4B-Safety-ThinkingI1-Q4_K_M4.2B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Huihui-Qwen3.5-4B-abliteratedI1-Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Darkidol-Ballad-4BI1-Q4_K_M4.5B2.52 GiB4.00 GiB7.43 GiB0.01 GiB27±26.5%
Gemma-3-4B-VL-it-Gemini-Pro-Heretic-Uncensored-ThinkingQ8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
gemma-3-4b-it-roleplay-tuned-v1Q8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
medgemma-1.5-4b-itQ8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
gemma-3-4b-it-roleplay-tuned-v2Q8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
medgemma-4b-itQ8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
gemma-3-4b-it-heretic-uncensored-abliterated-ExtremeQ8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
gemma-3-4b-it-abliteratedQ8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
amoral-gemma3-4B-v1Q8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
Gemma3-4B-CodeCenturionQ8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
gemma-3-4b-itQ8_04.3B3.85 GiB2.67 GiB7.43 GiB0.01 GiB27±26.5%
Dolphin3.0-Qwen2.5-3bQ5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Qwen2.5-Coder-3B-Instruct-abliteratedI1-Q5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
GRM-Kerlin-3b-AbliteratedI1-Q5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Qwen2.5-Coder-3B-InstructQ5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Mythos-nanoI1-Q5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
MATE-3BI1-Q5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Mythos-nano-OBLITERATEDI1-Q5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Qwen2.5-3B-Instruct-UncensoredI1-Q5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Nanonets-OCR-sQ5_K_S3.8B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Qwen2.5-3B-InstructQ5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Qwen2.5-Coder-3BQ5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
raspberry-3BQ5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
VibeThinker-3B-OBLITERATEDI1-Q5_K_S3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
VibeThinker-3BQ5_03.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Fourier-Qwen2.5-VL-3B-0.67I1-Q5_K_S3.8B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
Qwen2.5-VL-3B-InstructQ5_K_S3.8B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
jina-embeddings-v4Q5_K_S3.8B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
granite-3.1-1b-a400m-instructMoEIQ3_M1.3B0.57 GiB6.00 GiB7.43 GiB0.01 GiB16±37%
Qwen2.5-3B-Instruct-abliteratedI1-IQ2_XS3.1B2.02 GiB4.50 GiB7.43 GiB0.01 GiB27±26.5%
gemma-4-E4B-it-hereticQ4_18.0B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
gemma-4-E4B-uncensoredI1-IQ4_XS7.9B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
gemma-4-E4B-it-qat-q4_0-unquantized-hereticI1-IQ4_XS7.9B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
gemma-4-E4B-it-qat-heretic_decensoredI1-IQ4_XS7.9B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
gemma-4-E4B-it-QAT-SOMPOA-heresyI1-IQ4_XS7.9B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
gemma4-e4b-mahou-nsfwI1-IQ4_XS7.9B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
gemma-4-E4B-it-mentalchat16kI1-IQ4_XS7.9B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
gemma4-E4B-it-abliteratedI1-IQ4_XS7.9B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
gemma-4-E4B-it-OBLITERATEDI1-IQ4_XS8.0B4.69 GiB1.82 GiB7.43 GiB0.01 GiB27±26.5%
Gemma-4-E4B-LuchadorQ3_K_L8.0B4.69 GiB1.82 GiB7.42 GiB0.02 GiB27±26.5%
Holo-3.1-4BI1-Q3_K_L5.2B2.51 GiB4.00 GiB7.42 GiB0.02 GiB27±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 generation3.97 it/s2.63–4.85138
Benchmarked· n=138

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 6650 XT run?
431 of 2118 indexed open-weight models fit a Radeon RX 6650 XT at 131,072 context with f16 KV cache, the largest being Agents-A1-4B-Heretic-ARA-Refusals8 at Q4_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon RX 6650 XT actually have?
Its nameplate is 8 GB, but about 7.44 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon RX 6650 XT fast for local AI?
Its memory bandwidth is 280 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.