AMD · workstation

Radeon AI Pro R9700

Radeon AI Pro R9700 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 4K 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
300 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
vision language 181text 1739video 16image 2embedding 26audio tts 21audio asr 39

What fits at 4K context

largest quantization that fits, per model · 2024 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3.5-35B-A3BMoEQ6_K36.0B28.82 GiB0.02 GiB29.75 GiB0.01 GiB80±37%
Qwen3.6-35B-A3BMoEQ6_K36.0B28.82 GiB0.02 GiB29.75 GiB0.01 GiB80±37%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEIQ3_XXS—28.78 GiB0.03 GiB29.70 GiB0.06 GiB90±37%
Llama-3_1-Nemotron-51B-InstructIQ4_XS51.5B25.83 GiB2.81 GiB29.69 GiB0.07 GiB14±26.5%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.66 GiB0.10 GiB14±26.5%
Hypernova-60B-2605MoEI1-IQ3_S58.7B28.73 GiB0.04 GiB29.66 GiB0.10 GiB68±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-IQ3_XXS79.7B28.68 GiB0.03 GiB29.60 GiB0.16 GiB90±37%
Salience-1.5-ProMoEQ6_K_L36.0B28.66 GiB0.02 GiB29.59 GiB0.17 GiB81±37%
Qwable-v1MoEQ6_K_L36.0B28.66 GiB0.02 GiB29.59 GiB0.17 GiB81±37%
T-SearchMoEQ6_K_L36.0B28.66 GiB0.02 GiB29.59 GiB0.17 GiB81±37%
L3-DARKEST-PLANET-16.5BIQ4_XS16.5B28.30 GiB0.31 GiB29.55 GiB0.21 GiB14±26.5%
Melody1437-27BQ3_K_M27.8B28.40 GiB0.07 GiB29.44 GiB0.32 GiB14±26.5%
uyu-2-28BQ8_028.2B27.92 GiB0.51 GiB29.41 GiB0.35 GiB14±26.5%
Qwen2.5-7B-Instruct-1MF327.6B28.38 GiB0.06 GiB29.39 GiB0.37 GiB14±26.5%
DeepSeek-R1-Distill-Qwen-7BF327.6B28.38 GiB0.06 GiB29.39 GiB0.37 GiB14±26.5%
UI-TARS-7B-DPOF328.3B28.38 GiB0.06 GiB29.39 GiB0.37 GiB14±26.5%
Qwen2-7B-InstructF327.6B28.38 GiB0.06 GiB29.39 GiB0.37 GiB14±26.5%
Hercules-5.0-Qwen2-7BF327.6B28.38 GiB0.06 GiB29.39 GiB0.37 GiB14±26.5%
Kepler-8B-Instruct-v2F167.6B28.37 GiB0.06 GiB29.39 GiB0.37 GiB14±26.5%
MiniCPM-o-2_6F328.7B28.37 GiB0.06 GiB29.38 GiB0.38 GiB14±26.5%
Gemma-3-27B-MeditronFOQ8_028.8B28.13 GiB0.26 GiB29.36 GiB0.40 GiB14±26.5%
CalmeRys-78B-Orpo-v0.1I1-IQ2_S78.0B27.87 GiB0.38 GiB29.28 GiB0.48 GiB14±26.5%
Seed-OSS-36B-InstructQ6_K_L36.2B27.99 GiB0.28 GiB29.27 GiB0.49 GiB14±26.5%
Hermes-4.3-36BQ6_K_L36.2B27.99 GiB0.28 GiB29.27 GiB0.49 GiB14±26.5%
Qwen3.5-88BMoEI1-Q2_K_S87.7B28.30 GiB0.03 GiB29.25 GiB0.51 GiB73±37%
CodeLlama-70b-Instruct-hfI1-IQ3_S69.0B27.86 GiB0.35 GiB29.24 GiB0.52 GiB14±26.5%
CodeLlama-70b-Python-hfI1-IQ3_S69.0B27.86 GiB0.35 GiB29.24 GiB0.52 GiB14±26.5%
Nous-Hermes-Llama2-70bI1-IQ3_S69.0B27.86 GiB0.35 GiB29.24 GiB0.52 GiB14±26.5%
Midnight-Miqu-70B-v1.5I1-IQ3_S69.0B27.86 GiB0.35 GiB29.24 GiB0.52 GiB14±26.5%
KafkaLM-70B-German-V0.1Q3_K_S69.0B27.86 GiB0.35 GiB29.24 GiB0.52 GiB14±26.5%
llama2_70b_chat_uncensoredQ3_K_S69.0B27.86 GiB0.35 GiB29.24 GiB0.52 GiB14±26.5%
Xwin-LM-70b-V0.1Q3_K_S69.0B27.86 GiB0.35 GiB29.24 GiB0.52 GiB14±26.5%
Llama-2-70b-chat-hfQ3_K_S69.0B27.86 GiB0.35 GiB29.24 GiB0.52 GiB14±26.5%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ2_XXS109B28.09 GiB0.21 GiB29.23 GiB0.53 GiB61±37%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q4_K_S53.0B28.13 GiB0.18 GiB29.21 GiB0.55 GiB55±37%
Kimi-Linear-48B-A3B-InstructMoEQ4_K_L49.1B28.26 GiB0.03 GiB29.20 GiB0.56 GiB14±26.5%
command-r-35b-writer-v2I1-Q6_K35.0B26.74 GiB1.41 GiB29.15 GiB0.61 GiB15±26.5%
Rombo-LLM-V3.0-Qwen-72bI1-Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Qwen2.5-72B-Instruct-abliteratedI1-Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Qwen2.5-72B-Instruct-abliterated-v2I1-Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
HuatuoGPT-o1-72BQ2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
MiroThinker-v1.0-72BI1-Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
EVA-Qwen2.5-72B-v0.2Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Qwen2.5-Math-72B-InstructQ2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Qwen2.5-72B-InstructQ2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Malaysian-Qwen2.5-72B-InstructI1-Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Qwen2.5-72BI1-Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
magnum-v4-72bI1-Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
KAT-Dev-72B-ExpQ2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Homer-v1.0-Qwen2.5-72BQ2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Qwen2.5-VL-72B-InstructQ2_K73.4B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Chuluun-Qwen2.5-72B-v0.01Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Chronos-Platinum-72BQ2_K72.7B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
UI-TARS-72B-DPOQ2_K73.4B27.76 GiB0.35 GiB29.15 GiB0.61 GiB15±26.5%
Darwin-35B-A3B-OpusMoEQ6_K_L36.0B28.22 GiB0.02 GiB29.15 GiB0.61 GiB82±37%
Aurora-Code-1MoEQ6_K_L34.7B28.22 GiB0.02 GiB29.15 GiB0.61 GiB82±37%
grug-35b-v2MoEQ6_K_L35.1B28.22 GiB0.02 GiB29.15 GiB0.61 GiB82±37%
grug-35bMoEQ6_K_L35.1B28.22 GiB0.02 GiB29.15 GiB0.61 GiB82±37%
WorldSim-Opus-3.6-35B-A3BMoEQ6_K_L35.1B28.22 GiB0.02 GiB29.15 GiB0.61 GiB82±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 R9700 run?
2024 of 2118 indexed open-weight models fit a Radeon AI Pro R9700 at 4,096 context with q4_0 KV cache, the largest being Qwen3.5-35B-A3B at Q6_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon AI Pro R9700 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 R9700 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.