AMD · workstation

Radeon AI Pro R9600D

Radeon AI Pro R9600D has 32 GB of VRAM at 640 GB/s — about 29.76 GiB usable after driver and compositor overhead. 1998 of 2118 indexed models fit at 16K 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 1714video 16vision language 180image 2embedding 26audio asr 39audio tts 21

What fits at 16K context

largest quantization that fits, per model · 1998 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Hunyuan-A13B-InstructMoEUD-IQ2_M80.4B26.85 GiB2.00 GiB29.75 GiB0.01 GiB14±26.5%
Phi-3.5-MoE-instructMoEKV unresolvedQ5_K_S41.9B26.84 GiB2.00 GiB29.74 GiB0.02 GiB33±37%
CodeLlama-70b-Instruct-hfI1-Q2_K69.0B23.71 GiB5.00 GiB29.74 GiB0.02 GiB14±26.5%
CodeLlama-70b-Python-hfI1-Q2_K69.0B23.71 GiB5.00 GiB29.74 GiB0.02 GiB14±26.5%
Nous-Hermes-Llama2-70bI1-Q2_K69.0B23.71 GiB5.00 GiB29.74 GiB0.02 GiB14±26.5%
Midnight-Miqu-70B-v1.5I1-Q2_K69.0B23.71 GiB5.00 GiB29.74 GiB0.02 GiB14±26.5%
KafkaLM-70B-German-V0.1Q2_K69.0B23.71 GiB5.00 GiB29.74 GiB0.02 GiB14±26.5%
GPT-NeoX-20B-ErebusI1-Q4_K_M20.6B12.23 GiB16.50 GiB29.73 GiB0.03 GiB14±26.5%
Assistant_Pepe_70BIQ2_S70.6B23.67 GiB5.00 GiB29.69 GiB0.07 GiB14±26.5%
Mixtral_34Bx2_MoE_60BMoEQ3_K_M60.8B24.95 GiB3.75 GiB29.68 GiB0.08 GiB8±37%
Qwen3-Coder-NextMoEQ2_K_L79.7B27.29 GiB1.50 GiB29.67 GiB0.09 GiB60±37%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.66 GiB0.10 GiB14±26.5%
Salience-1.5-ProMoEQ6_K36.0B28.43 GiB0.31 GiB29.65 GiB0.11 GiB74±37%
Qwable-v1MoEQ6_K36.0B28.43 GiB0.31 GiB29.65 GiB0.11 GiB74±37%
T-SearchMoEQ6_K36.0B28.43 GiB0.31 GiB29.65 GiB0.11 GiB74±37%
Kimi-Linear-48B-A3B-InstructMoEQ4_K_L49.1B28.26 GiB0.47 GiB29.64 GiB0.12 GiB14±26.5%
OLMo-2-0325-32BQ6_K32.2B24.63 GiB4.00 GiB29.63 GiB0.13 GiB14±26.5%
Qwen3-Next-80B-A3B-ThinkingMoEQ2_K_L81.3B27.24 GiB1.50 GiB29.63 GiB0.13 GiB60±37%
Qwen3-Next-80B-A3B-InstructMoEQ2_K_L81.3B27.24 GiB1.50 GiB29.63 GiB0.13 GiB60±37%
Qwen3.5-88BMoEI1-Q2_K_S87.7B28.30 GiB0.38 GiB29.60 GiB0.16 GiB66±37%
medgemma-27b-itQ8_028.8B26.74 GiB1.86 GiB29.58 GiB0.18 GiB14±26.5%
gemma-3-27b-it-abliterated-refined-visionQ8_027.4B26.74 GiB1.86 GiB29.58 GiB0.18 GiB14±26.5%
gemma-3-27b-it-abliteratedQ8_027.4B26.74 GiB1.86 GiB29.58 GiB0.18 GiB14±26.5%
Nidum-Gemma-3-27B-it-UncensoredQ8_027.4B26.74 GiB1.86 GiB29.58 GiB0.18 GiB14±26.5%
gemma-3-27b-itQ8_027.4B26.74 GiB1.86 GiB29.58 GiB0.18 GiB14±26.5%
Unbound-v1.12.0-27BQ8_027.4B26.74 GiB1.86 GiB29.58 GiB0.18 GiB14±26.5%
medgemma-27b-text-itQ8_027.0B26.74 GiB1.86 GiB29.58 GiB0.18 GiB14±26.5%
Gemma4-Gutenberg-31BQ6_K31.3B24.89 GiB3.67 GiB29.54 GiB0.22 GiB14±26.5%
gemma-4-31B-itQ6_K31.3B24.89 GiB3.67 GiB29.54 GiB0.22 GiB14±26.5%
Gemma4-Gutenberg-31B-HereticQ6_K31.3B24.89 GiB3.67 GiB29.54 GiB0.22 GiB14±26.5%
Equinox-31BQ6_K31.3B24.89 GiB3.67 GiB29.54 GiB0.22 GiB14±26.5%
gemma-4-31B-it-SDFT-Heretic-RPQ6_K30.7B24.89 GiB3.67 GiB29.54 GiB0.22 GiB14±26.5%
command-r-35b-writer-v2I1-IQ1_M35.0B8.52 GiB20.00 GiB29.52 GiB0.24 GiB14±26.5%
xLAM-8x7b-rMoEQ4_K_L46.7B26.59 GiB2.00 GiB29.52 GiB0.24 GiB23±37%
Darwin-35B-A3B-OpusMoEQ6_K_L36.0B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
Aurora-Code-1MoEQ6_K_L34.7B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
grug-35b-v2MoEQ6_K_L35.1B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
grug-35bMoEQ6_K_L35.1B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
WorldSim-Opus-3.6-35B-A3BMoEQ6_K_L35.1B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
Qwen3.6-35B-A3B-AnkoMoEQ6_K_L35.1B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
KAT-Coder-V2.5-DevMoEQ6_K_L34.7B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
Ornith-1.0-35BMoEQ6_K_L34.7B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
Nex-N2-miniMoEQ6_K_L35.1B28.22 GiB0.31 GiB29.44 GiB0.32 GiB75±37%
dolphin-2.6-mixtral-8x7bMoEI1-Q4_K_M46.7B26.49 GiB2.00 GiB29.43 GiB0.33 GiB23±37%
Nous-Hermes-2-Mixtral-8x7B-DPOMoEQ4_K_M46.7B26.49 GiB2.00 GiB29.43 GiB0.33 GiB23±37%
Mixtral-8x7B-Instruct-v0.1MoEQ4_K_M46.7B26.49 GiB2.00 GiB29.43 GiB0.33 GiB23±37%
dolphin-2.5-mixtral-8x7bMoEQ4_K_M46.7B26.49 GiB2.00 GiB29.43 GiB0.33 GiB23±37%
Mixtral-8x7B-v0.1MoEQ4_K_M46.7B26.49 GiB2.00 GiB29.43 GiB0.33 GiB23±37%
Qwen3.6-27B-Fable-5-ExperimentalQ8_027.8B27.42 GiB1.00 GiB29.39 GiB0.37 GiB14±26.5%
Open_Gpt4_8x7B_v0.2MoEQ4_K_M46.7B26.43 GiB2.00 GiB29.37 GiB0.39 GiB23±37%
Noromaid-20b-v0.1.1I1-Q3_K_M20.0B9.04 GiB19.38 GiB29.36 GiB0.40 GiB14±26.5%
Nethena-20BQ3_K_M20.0B9.03 GiB19.38 GiB29.35 GiB0.41 GiB14±26.5%
Magistral-Small-2509-VisionQ6_K_L24.0B25.83 GiB2.50 GiB29.34 GiB0.42 GiB14±26.5%
Hypernova-60B-2605MoEI1-IQ3_XXS58.7B27.90 GiB0.52 GiB29.31 GiB0.45 GiB62±37%
Seed-OSS-36B-InstructQ5_K_L36.2B24.29 GiB4.00 GiB29.29 GiB0.47 GiB14±26.5%
Hermes-4.3-36BQ5_K_L36.2B24.29 GiB4.00 GiB29.29 GiB0.47 GiB14±26.5%
IQuest-Coder-V1-40B-InstructI1-Q4_139.8B23.24 GiB5.00 GiB29.24 GiB0.52 GiB14±26.5%
gemma-4-E4B-it-Uncensored-MAXF328.0B28.02 GiB0.29 GiB29.23 GiB0.53 GiB14±26.5%
Qwen3.6-28BMoEQ8_028.2B28.00 GiB0.31 GiB29.22 GiB0.54 GiB70±37%
Apertus-70B-Instruct-2509UD-IQ2_M70.6B23.12 GiB5.00 GiB29.20 GiB0.56 GiB15±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 R9600D run?
1998 of 2118 indexed open-weight models fit a Radeon AI Pro R9600D at 16,384 context with f16 KV cache, the largest being Hunyuan-A13B-Instruct at UD-IQ2_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon AI Pro R9600D 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 R9600D 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.