AMD · workstation

Radeon Pro W6600

Radeon Pro W6600 has 8 GB of VRAM at 224 GB/s — about 7.44 GiB usable after driver and compositor overhead. 1413 of 2118 indexed models fit at 4K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
8 GB
GDDR6
Bandwidth
224 GB/s
128-bit bus
Tensor FP16
—
dense
TDP
130 W
$649 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1217vision language 102embedding 26audio tts 21video 7image 2audio asr 38

What fits at 4K context

largest quantization that fits, per model · 1413 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Ling-liteMoEQ2_K16.8B6.43 GiB0.12 GiB7.44 GiB0.00 GiB70±37%
granite-4.1-8bQ5_18.8B6.17 GiB0.33 GiB7.44 GiB0.00 GiB22±26.5%
Phi-4-reasoningQ3_K_S14.7B6.06 GiB0.42 GiB7.43 GiB0.01 GiB22±26.5%
Phi-4-reasoning-plusQ3_K_S14.7B6.06 GiB0.42 GiB7.43 GiB0.01 GiB22±26.5%
phi-4Q3_K_S14.7B6.06 GiB0.42 GiB7.43 GiB0.01 GiB22±26.5%
Falcon3-10B-InstructQ4_K_L10.3B6.13 GiB0.33 GiB7.43 GiB0.01 GiB22±26.5%
gemma-3-12b-it-vl-Gemini-3-Pro-Preview-Heretic-Uncensored-ThinkingI1-IQ4_XS12.2B6.10 GiB0.38 GiB7.43 GiB0.01 GiB22±26.5%
gemma-3-12b-it-vl-Deepseek-v3.1-Heretic-Uncensored-ThinkingI1-IQ4_XS12.2B6.10 GiB0.38 GiB7.43 GiB0.01 GiB22±26.5%
gemma-3-12b-it-vl-GLM-4.7-Flash-Heretic-Uncensored-ThinkingI1-IQ4_XS12.2B6.10 GiB0.38 GiB7.43 GiB0.01 GiB22±26.5%
Floppa-12B-Gemma3-UncensoredI1-IQ4_XS12.2B6.10 GiB0.38 GiB7.43 GiB0.01 GiB22±26.5%
gemma-3-12b-it-hereticI1-IQ4_XS12.2B6.10 GiB0.38 GiB7.43 GiB0.01 GiB22±26.5%
gemma-3-12b-it-abliteratedIQ4_XS12.2B6.10 GiB0.38 GiB7.43 GiB0.01 GiB22±26.5%
gemma-3-12b-itIQ4_XS12.2B6.10 GiB0.38 GiB7.43 GiB0.01 GiB22±26.5%
DeepCoder-14B-PreviewQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
SuperNova-MediusQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
Qwen2.5-Coder-14B-Instruct-abliteratedQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
OpenCodeReasoning-Nemotron-14BQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
0x-liteQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
Qwen2.5-Coder-14B-InstructQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
Qwen2.5-14B-InstructQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
Qwen2.5-14B-Instruct-1MQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
Qwen2.5-Coder-14BQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
DeepSeek-R1-Distill-Qwen-14BQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
AceReason-Nemotron-14BQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
UwU-14B-Math-v0.2Q2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
oxy-1-smallQ2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
EVA-Qwen2.5-14B-v0.2Q2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
EVA-Qwen2.5-14B-v0.0Q2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
EVA-Qwen2.5-14B-v0.1Q2_K_L14.8B6.08 GiB0.40 GiB7.43 GiB0.01 GiB22±26.5%
MythoMax-L2-13bI1-IQ3_XXS13.0B4.82 GiB1.66 GiB7.42 GiB0.02 GiB22±26.5%
Lamarck-14B-v0.7Q2_K_L14.8B6.08 GiB0.40 GiB7.42 GiB0.02 GiB22±26.5%
NVIDIA-Nemotron-Nano-12B-v2Q3_K_L12.3B5.94 GiB0.51 GiB7.42 GiB0.02 GiB22±26.5%
Wan2.1-T2V-1.3BQ4_01.4B6.50 GiB0.00 GiB7.42 GiB0.02 GiB22±26.5%
Rocinante-XL-16B-v1Q2_K16.1B6.02 GiB0.45 GiB7.42 GiB0.02 GiB22±26.5%
Qwen2.5-3B-Instruct-abliteratedQ8_03.1B6.43 GiB0.07 GiB7.41 GiB0.03 GiB22±26.5%
glm-4-9b-chat-1mQ4_09.5B5.14 GiB1.33 GiB7.41 GiB0.03 GiB22±26.5%
GLM-4.7-Flash-DerestrictedMoEI1-IQ1_M31.2B6.39 GiB0.11 GiB7.41 GiB0.03 GiB84±37%
Huihui-GLM-4.7-Flash-abliteratedMoEI1-IQ1_M31.2B6.39 GiB0.11 GiB7.41 GiB0.03 GiB84±37%
granite-vision-4.1-4bBF164.0B6.34 GiB0.17 GiB7.41 GiB0.03 GiB22±26.5%
Parable-Granite-4.1-3B-Claude-Fable-5F163.4B6.34 GiB0.17 GiB7.41 GiB0.03 GiB22±26.5%
granite-4.0-microBF163.4B6.34 GiB0.17 GiB7.41 GiB0.03 GiB22±26.5%
granite-4.1-3bBF163.4B6.34 GiB0.17 GiB7.41 GiB0.03 GiB22±26.5%
granite-4.0-micro-baseBF163.4B6.34 GiB0.17 GiB7.41 GiB0.03 GiB22±26.5%
dolphincoder-starcoder2-15bKV unresolvedI1-IQ3_XS16.0B6.25 GiB0.17 GiB7.41 GiB0.03 GiB22±26.5%
starcoder2-15bKV unresolvedIQ3_XS16.0B6.25 GiB0.17 GiB7.41 GiB0.03 GiB22±26.5%
LFM2-24B-A2BMoEIQ2_S23.8B6.45 GiB0.04 GiB7.41 GiB0.03 GiB89±37%
Apriel-1.6-15b-ThinkerI1-IQ3_S14.9B6.06 GiB0.40 GiB7.40 GiB0.04 GiB22±26.5%
LFM2.5-Queen-Opus-4.7-8B-A1BMoEI1-Q6_K8.5B6.48 GiB0.02 GiB7.40 GiB0.04 GiB64±37%
LFM2.5-8B-A1B-KO-SFTMoEI1-Q6_K8.5B6.48 GiB0.02 GiB7.40 GiB0.04 GiB64±37%
LFM2.5-8B-A1B-hereticMoEI1-Q6_K8.5B6.48 GiB0.02 GiB7.40 GiB0.04 GiB64±37%
LFM2.5-8B-A1B-SOMPOA-heresyMoEI1-Q6_K8.5B6.48 GiB0.02 GiB7.40 GiB0.04 GiB64±37%
Huihui-LFM2.5-8B-A1B-abliteratedMoEI1-Q6_K8.5B6.48 GiB0.02 GiB7.40 GiB0.04 GiB64±37%
Supertron2.1-8B-A1BMoEI1-Q6_K8.5B6.48 GiB0.02 GiB7.40 GiB0.04 GiB64±37%
LFM2.5-8B-A1BMoEQ6_K8.5B6.48 GiB0.02 GiB7.40 GiB0.04 GiB64±37%
Phi-3-medium-128k-instructIQ3_M14.0B6.03 GiB0.42 GiB7.40 GiB0.04 GiB22±26.5%
Phi-3-medium-4k-instructI1-IQ3_M14.0B6.03 GiB0.42 GiB7.40 GiB0.04 GiB22±26.5%
glm-4-9b-chatIQ4_NL9.4B5.13 GiB1.33 GiB7.40 GiB0.04 GiB22±26.5%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEI1-IQ2_XS23.0B6.38 GiB0.11 GiB7.40 GiB0.04 GiB75±37%
reka-flash-3.1I1-IQ1_S20.9B6.15 GiB0.27 GiB7.40 GiB0.04 GiB22±26.5%
Gemma-4-12B-StyleTuneI1-Q3_K_M13.0B6.07 GiB0.38 GiB7.40 GiB0.04 GiB22±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 Pro W6600 run?
1413 of 2118 indexed open-weight models fit a Radeon Pro W6600 at 4,096 context with q8_0 KV cache, the largest being Ling-lite at Q2_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon Pro W6600 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 Pro W6600 fast for local AI?
Its memory bandwidth is 224 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.