AMD · workstation

Radeon AI Pro R9600

Radeon AI Pro R9600 has 32 GB of VRAM at 640 GB/s — about 29.76 GiB usable after driver and compositor overhead. 1863 of 2118 indexed models fit at 64K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
32 GB
GDDR6
Bandwidth
640 GB/s
256-bit bus
Tensor FP16
—
dense
TDP
150 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1584vision language 176video 16audio asr 39audio tts 21image 1embedding 26

What fits at 64K context

largest quantization that fits, per model · 1863 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
OpenBuddy-R1-0528-Distill-Qwen3-32B-Preview0-QATIQ3_XS32.8B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
Qwen3-VL-32B-Instruct-ultra-uncensored-hereticI1-IQ3_XS33.4B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
Huihui-Qwen3-VL-32B-Instruct-abliteratedI1-IQ3_XS33.4B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
KAT-DevIQ3_XS32.8B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
ColorGUI-32BI1-IQ3_XS33.4B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
Qwen3-VL-32B-InstructIQ3_XS33.4B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
Qwen3-32B-UncensoredI1-IQ3_XS32.8B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
Qwen3-32B-abliteratedI1-IQ3_XS32.8B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
DeepSWE-PreviewIQ3_XS32.8B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
AReaL-boba-2-32BI1-IQ3_XS32.8B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
Assistant_Pepe_32BI1-IQ3_XS32.8B12.76 GiB16.00 GiB29.75 GiB0.01 GiB14±26.5%
Phi-3-mini-4k-instructKV unresolvedQ2_K3.8B4.85 GiB24.00 GiB29.75 GiB0.01 GiB14±26.5%
Nous-Capybara-limarpv3-34BI1-IQ3_XS34.4B13.76 GiB15.00 GiB29.75 GiB0.01 GiB14±26.5%
GRM-2.6-Plus-0628Q6_K_M27.8B24.76 GiB4.00 GiB29.73 GiB0.03 GiB14±26.5%
Gemma-4-Novelist-Eclipse-31BQ4_032.7B17.57 GiB11.17 GiB29.73 GiB0.03 GiB14±26.5%
Gemma-4-31B-StyleTuneQ4_032.7B17.57 GiB11.17 GiB29.73 GiB0.03 GiB14±26.5%
Trinity-2-Codestral-22B-v0.2Q5_K_L22.2B14.76 GiB14.00 GiB29.72 GiB0.04 GiB14±26.5%
Mistral-Small-Drummer-22BQ5_K_L22.2B14.76 GiB14.00 GiB29.72 GiB0.04 GiB14±26.5%
Mistral-Small-Instruct-2409Q5_K_L22.2B14.76 GiB14.00 GiB29.72 GiB0.04 GiB14±26.5%
Mistral-Small-22B-ArliAI-RPMax-v1.1Q5_K_L22.2B14.76 GiB14.00 GiB29.72 GiB0.04 GiB14±26.5%
magnum-v4-22bQ5_K_L22.2B14.76 GiB14.00 GiB29.72 GiB0.04 GiB14±26.5%
GLM-4-32B-0414-Korean-CultureI1-Q6_K32.6B24.89 GiB3.81 GiB29.69 GiB0.07 GiB14±26.5%
GLM-Z1-32B-0414Q6_K32.6B24.89 GiB3.81 GiB29.69 GiB0.07 GiB14±26.5%
GLM-4-32B-0414Q6_K32.6B24.89 GiB3.81 GiB29.69 GiB0.07 GiB14±26.5%
GLM-Z1-32B-0414-uncensored-heretic-v2Q6_K32.6B24.89 GiB3.81 GiB29.69 GiB0.07 GiB14±26.5%
Magistry-24B-v1.1Q6_K_L23.6B18.67 GiB10.00 GiB29.69 GiB0.07 GiB14±26.5%
Yi-34B-200K-DARE-megamerge-v8IQ3_XS34.4B13.71 GiB15.00 GiB29.69 GiB0.07 GiB14±26.5%
Nous-Hermes-2-Yi-34BI1-IQ3_XS34.4B13.71 GiB15.00 GiB29.69 GiB0.07 GiB14±26.5%
Phi-3.5-MoE-instructMoEKV unresolvedIQ4_XS41.9B20.78 GiB8.00 GiB29.69 GiB0.07 GiB20±37%
Qwen2.5-Coder-14B-InstructQ4_K_M14.8B16.74 GiB12.00 GiB29.69 GiB0.07 GiB14±26.5%
IQuest-Coder-V1-40B-InstructI1-IQ1_M39.8B8.68 GiB20.00 GiB29.68 GiB0.08 GiB14±26.5%
Seed-OSS-36B-Instruct-biprojected-norm-preserving-abliteratedI1-Q2_K36.2B12.67 GiB16.00 GiB29.67 GiB0.09 GiB14±26.5%
Seed-OSS-36B-InstructQ2_K36.2B12.67 GiB16.00 GiB29.67 GiB0.09 GiB14±26.5%
Hermes-4.3-36B-hereticI1-Q2_K36.2B12.67 GiB16.00 GiB29.67 GiB0.09 GiB14±26.5%
Hermes-4.3-36BQ2_K36.2B12.67 GiB16.00 GiB29.67 GiB0.09 GiB14±26.5%
magnum-v2-32bIQ3_XS32.5B12.67 GiB16.00 GiB29.67 GiB0.09 GiB14±26.5%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.66 GiB0.10 GiB14±26.5%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ2_K—27.26 GiB1.50 GiB29.65 GiB0.11 GiB60±37%
Qwen3-Next-80B-A3B-InstructMoEUD-IQ1_M81.3B22.73 GiB6.00 GiB29.62 GiB0.14 GiB29±37%
Cydonia-v1.3-Magnum-v4-22BI1-Q5_K_M22.2B14.64 GiB14.00 GiB29.60 GiB0.16 GiB14±26.5%
Codestral-22B-v0.1-hfQ5_K_M22.2B14.64 GiB14.00 GiB29.60 GiB0.16 GiB14±26.5%
Codestral-22B-v0.1Q5_K_M22.2B14.64 GiB14.00 GiB29.60 GiB0.16 GiB14±26.5%
dolphin-2.9.1-mixtral-1x22bMoEI1-Q5_K_M22.2B14.64 GiB14.00 GiB29.60 GiB0.16 GiB8±37%
Voxtral-Small-24B-2507Q6_K24.3B18.57 GiB10.00 GiB29.59 GiB0.17 GiB14±26.5%
gemma-4-26B-A4B-itMoEQ8_026.5B25.89 GiB2.79 GiB29.57 GiB0.19 GiB14±26.5%
Gemma-4-Gembrain-X-Core-31BI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Gemma-4-Gembrain-X-31BI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Gemma-4-31B-Isometry-Fabled-PersonaI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Versipellis-31BI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Gemma4-Gutenberg-31BI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
G4-MeroMero-31B-uncensored-hereticI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Gemma-4-Novelist-31BI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Wanabi-Gemma4-31BI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
G4-Alice-v1.2-31BI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Agares-31B-v1I1-Q4_K_M30.7B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Gemma4-Gutenberg-31B-HereticI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-hereticI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Gemma-4-Gemsicle-31BI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Gemma-4-Gembrain-31B-it-uncensored-hereticI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±26.5%
Melinoe-Gemma4-31B-VL-hereticI1-Q4_K_M31.3B17.40 GiB11.17 GiB29.56 GiB0.20 GiB14±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 AI Pro R9600 run?
1863 of 2118 indexed open-weight models fit a Radeon AI Pro R9600 at 65,536 context with f16 KV cache, the largest being OpenBuddy-R1-0528-Distill-Qwen3-32B-Preview0-QAT at IQ3_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon AI Pro R9600 actually have?
Its nameplate is 32 GB, but about 29.76 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon AI Pro R9600 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.