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. 2024 of 2118 indexed models fit at 8K context with q4_0 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 1739video 16vision language 181image 2embedding 26audio tts 21audio asr 39

What fits at 8K context

largest quantization that fits, per model · 2024 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEIQ3_XXS—28.78 GiB0.05 GiB29.73 GiB0.03 GiB89±37%
v6-Finch-14B-HFF1614.1B26.63 GiB2.14 GiB29.72 GiB0.04 GiB14±26.5%
Hypernova-60B-2605MoEI1-IQ3_S58.7B28.73 GiB0.08 GiB29.69 GiB0.07 GiB68±37%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.66 GiB0.10 GiB14±26.5%
CalmeRys-78B-Orpo-v0.1I1-IQ2_S78.0B27.87 GiB0.76 GiB29.66 GiB0.10 GiB14±26.5%
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-IQ3_XXS79.7B28.68 GiB0.05 GiB29.62 GiB0.14 GiB90±37%
Salience-1.5-ProMoEQ6_K_L36.0B28.66 GiB0.04 GiB29.61 GiB0.15 GiB80±37%
Qwable-v1MoEQ6_K_L36.0B28.66 GiB0.04 GiB29.61 GiB0.15 GiB80±37%
T-SearchMoEQ6_K_L36.0B28.66 GiB0.04 GiB29.61 GiB0.15 GiB80±37%
CodeLlama-70b-Instruct-hfI1-IQ3_S69.0B27.86 GiB0.70 GiB29.59 GiB0.17 GiB14±26.5%
CodeLlama-70b-Python-hfI1-IQ3_S69.0B27.86 GiB0.70 GiB29.59 GiB0.17 GiB14±26.5%
Nous-Hermes-Llama2-70bI1-IQ3_S69.0B27.86 GiB0.70 GiB29.59 GiB0.17 GiB14±26.5%
Midnight-Miqu-70B-v1.5I1-IQ3_S69.0B27.86 GiB0.70 GiB29.59 GiB0.17 GiB14±26.5%
KafkaLM-70B-German-V0.1Q3_K_S69.0B27.86 GiB0.70 GiB29.59 GiB0.17 GiB14±26.5%
llama2_70b_chat_uncensoredQ3_K_S69.0B27.86 GiB0.70 GiB29.59 GiB0.17 GiB14±26.5%
Xwin-LM-70b-V0.1Q3_K_S69.0B27.86 GiB0.70 GiB29.59 GiB0.17 GiB14±26.5%
Llama-2-70b-chat-hfQ3_K_S69.0B27.86 GiB0.70 GiB29.59 GiB0.17 GiB14±26.5%
uyu-2-28BQ8_028.2B27.92 GiB0.68 GiB29.59 GiB0.17 GiB14±26.5%
Seed-OSS-36B-InstructQ6_K_L36.2B27.99 GiB0.56 GiB29.55 GiB0.21 GiB14±26.5%
Hermes-4.3-36BQ6_K_L36.2B27.99 GiB0.56 GiB29.55 GiB0.21 GiB14±26.5%
Melody1437-27BQ3_K_M27.8B28.40 GiB0.14 GiB29.51 GiB0.25 GiB14±26.5%
Rombo-LLM-V3.0-Qwen-72bI1-Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Qwen2.5-72B-Instruct-abliteratedI1-Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Qwen2.5-72B-Instruct-abliterated-v2I1-Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
HuatuoGPT-o1-72BQ2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
MiroThinker-v1.0-72BI1-Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
EVA-Qwen2.5-72B-v0.2Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Qwen2.5-Math-72B-InstructQ2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Qwen2.5-72B-InstructQ2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Malaysian-Qwen2.5-72B-InstructI1-Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Qwen2.5-72BI1-Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
magnum-v4-72bI1-Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
KAT-Dev-72B-ExpQ2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Homer-v1.0-Qwen2.5-72BQ2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Qwen2.5-VL-72B-InstructQ2_K73.4B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Chuluun-Qwen2.5-72B-v0.01Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Chronos-Platinum-72BQ2_K72.7B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
UI-TARS-72B-DPOQ2_K73.4B27.76 GiB0.70 GiB29.50 GiB0.26 GiB14±26.5%
Qwen2.5-7B-Instruct-1MF327.6B28.38 GiB0.12 GiB29.45 GiB0.31 GiB14±26.5%
DeepSeek-R1-Distill-Qwen-7BF327.6B28.38 GiB0.12 GiB29.45 GiB0.31 GiB14±26.5%
UI-TARS-7B-DPOF328.3B28.38 GiB0.12 GiB29.45 GiB0.31 GiB14±26.5%
Qwen2-7B-InstructF327.6B28.38 GiB0.12 GiB29.45 GiB0.31 GiB14±26.5%
Hercules-5.0-Qwen2-7BF327.6B28.38 GiB0.12 GiB29.45 GiB0.31 GiB14±26.5%
Gemma-3-27B-MeditronFOQ8_028.8B28.13 GiB0.35 GiB29.45 GiB0.31 GiB14±26.5%
Kepler-8B-Instruct-v2F167.6B28.37 GiB0.12 GiB29.45 GiB0.31 GiB14±26.5%
MiniCPM-o-2_6F328.7B28.37 GiB0.12 GiB29.44 GiB0.32 GiB14±26.5%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ2_XXS109B28.09 GiB0.42 GiB29.44 GiB0.32 GiB58±37%
Devstral-2-123B-Instruct-2512IQ1_M125B27.59 GiB0.77 GiB29.42 GiB0.34 GiB14±26.5%
Mistral-Medium-3.5-128BI1-IQ1_M128B27.59 GiB0.77 GiB29.42 GiB0.34 GiB14±26.5%
XORTRON-NXTXPRTXXLI1-IQ1_M128B27.59 GiB0.77 GiB29.42 GiB0.34 GiB14±26.5%
Qwen3-72B-SynthesisQ2_K72.7B27.68 GiB0.70 GiB29.42 GiB0.34 GiB14±26.5%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q4_K_S53.0B28.13 GiB0.37 GiB29.40 GiB0.36 GiB53±37%
Apertus-70B-Instruct-2509IQ3_XS70.6B27.55 GiB0.70 GiB29.34 GiB0.42 GiB14±26.5%
Llama-3_3-Nemotron-Super-49B-v1_5Q3_K_M49.9B22.64 GiB5.63 GiB29.31 GiB0.45 GiB14±26.5%
Valkyrie-49B-v2.1I1-Q3_K_M49.9B22.64 GiB5.63 GiB29.31 GiB0.45 GiB14±26.5%
Llama-3_3-Nemotron-Super-49B-v1Q3_K_M49.9B22.64 GiB5.63 GiB29.31 GiB0.45 GiB14±26.5%
Kimi-Dev-72BUD-IQ2_M72.7B27.56 GiB0.70 GiB29.30 GiB0.46 GiB14±26.5%
Snowpiercer-15B-v4BF1615.0B27.90 GiB0.44 GiB29.29 GiB0.47 GiB14±26.5%
Qwen3.5-88BMoEI1-Q2_K_S87.7B28.30 GiB0.05 GiB29.28 GiB0.48 GiB72±37%
●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?
2024 of 2118 indexed open-weight models fit a Radeon AI Pro R9600D at 8,192 context with q4_0 KV cache, the largest being Qwen3-Coder-Next-Opus-4.6-Reasoning-Distilled at IQ3_XXS. 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.