AMD · workstation

Radeon Pro W7900

Radeon Pro W7900 has 48 GB of VRAM at 864 GB/s — about 44.64 GiB usable after driver and compositor overhead. 2024 of 2118 indexed models fit at 64K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
48 GB
GDDR6
Bandwidth
864 GB/s
384-bit bus
Tensor FP16
—
dense
TDP
295 W
$3999 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1736vision language 184image 2video 16audio tts 21audio asr 39embedding 26

What fits at 64K context

largest quantization that fits, per model · 2024 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
INTELLECT-1-InstructF3210.2B38.05 GiB5.58 GiB44.57 GiB0.07 GiB13±26.5%
Assistant_Pepe_70BQ3_K_M70.6B32.89 GiB10.63 GiB44.54 GiB0.10 GiB13±26.5%
GLM-4.6VMoEIQ2_XS108B37.49 GiB6.11 GiB44.53 GiB0.11 GiB29±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_079.7B42.78 GiB0.80 GiB44.47 GiB0.17 GiB69±37%
GPT-NeoX-20B-ErebusI1-IQ3_S20.6B8.35 GiB35.06 GiB44.41 GiB0.23 GiB13±26.5%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q2_K125B42.67 GiB0.80 GiB44.39 GiB0.25 GiB63±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ2_K123B42.66 GiB0.80 GiB44.39 GiB0.25 GiB63±37%
GLM-4.5-AirMoEUD-IQ1_M110B37.31 GiB6.11 GiB44.35 GiB0.29 GiB29±37%
internlm2-math-plus-20bBF1619.9B37.00 GiB6.38 GiB44.33 GiB0.31 GiB13±26.5%
reka-flash-3BF1620.9B38.94 GiB4.38 GiB44.30 GiB0.34 GiB13±26.5%
reka-flash-3.1BF1620.9B38.94 GiB4.38 GiB44.30 GiB0.34 GiB13±26.5%
Mistral-Small-Instruct-2409IQ3_XS22.2B35.84 GiB7.44 GiB44.24 GiB0.40 GiB13±26.5%
Qwen2.5-Coder-32B-InstructQ4_032.8B34.72 GiB8.50 GiB44.22 GiB0.42 GiB13±26.5%
Qwen3.5-88BMoEI1-Q3_K_L87.7B42.43 GiB0.80 GiB44.16 GiB0.48 GiB57±37%
CalmeRys-78B-Orpo-v0.1I1-IQ3_XXS78.0B31.70 GiB11.42 GiB44.15 GiB0.49 GiB13±26.5%
calme-2.3-rys-78bIQ3_XXS78.0B31.70 GiB11.42 GiB44.15 GiB0.49 GiB13±26.5%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-Q2_K109B36.85 GiB6.38 GiB44.15 GiB0.49 GiB29±37%
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16MoEQ4_K_S75.4B43.15 GiB0.00 GiB44.12 GiB0.52 GiB102±37%
Devstral-2-123B-Instruct-2512IQ2_XXS125B31.35 GiB11.69 GiB44.10 GiB0.54 GiB13±26.5%
Mistral-Medium-3.5-128BI1-IQ2_XXS128B31.35 GiB11.69 GiB44.10 GiB0.54 GiB13±26.5%
XORTRON-NXTXPRTXXLI1-IQ2_XXS128B31.35 GiB11.69 GiB44.10 GiB0.54 GiB13±26.5%
Delphi-25B-SimpleRL-MathI1-IQ2_S25.0B7.55 GiB35.56 GiB44.09 GiB0.55 GiB13±26.5%
Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingQ8_039.5B39.90 GiB3.19 GiB44.06 GiB0.58 GiB13±26.5%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ4_K_S—42.37 GiB0.80 GiB44.05 GiB0.59 GiB70±37%
deepseek-llm-67b-chatI1-Q3_K_M67.4B30.41 GiB12.62 GiB44.03 GiB0.61 GiB13±26.5%
deepseek-llm-67b-baseI1-Q3_K_M67.4B30.41 GiB12.62 GiB44.03 GiB0.61 GiB13±26.5%
openbuddy-deepseek-67b-v15.3-4kI1-Q3_K_M67.4B30.41 GiB12.62 GiB44.03 GiB0.61 GiB13±26.5%
Hunyuan-A13B-InstructMoEQ3_K_L80.4B38.84 GiB4.25 GiB43.99 GiB0.65 GiB13±26.5%
Qwen3-Coder-NextMoEIQ4_XS79.7B39.91 GiB3.19 GiB43.98 GiB0.66 GiB47±37%
Qwen3-Next-80B-A3B-ThinkingMoEIQ4_XS81.3B39.91 GiB3.19 GiB43.98 GiB0.66 GiB47±37%
Qwen3-Next-80B-A3B-InstructMoEIQ4_XS81.3B39.91 GiB3.19 GiB43.98 GiB0.66 GiB47±37%
GLM-4.5-Air-DerestrictedMoEIQ2_XXS110B36.90 GiB6.11 GiB43.94 GiB0.70 GiB29±37%
Laguna-S-2.1MoEUD-IQ3_XXS118B41.24 GiB1.67 GiB43.83 GiB0.81 GiB52±37%
Rombo-LLM-V3.0-Qwen-72bI1-IQ3_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Qwen2.5-72B-Instruct-abliteratedI1-IQ3_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ3_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
HuatuoGPT-o1-72BQ3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
MiroThinker-v1.0-72BI1-IQ3_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
EVA-Qwen2.5-72B-v0.2Q3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Qwen2.5-Math-72B-InstructQ3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Qwen2.5-72B-InstructQ3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Malaysian-Qwen2.5-72B-InstructI1-IQ3_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Qwen2.5-72BI1-IQ3_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
magnum-v4-72bI1-IQ3_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Kimi-Dev-72BQ3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
KAT-Dev-72B-ExpQ3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Homer-v1.0-Qwen2.5-72BQ3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Chuluun-Qwen2.5-72B-v0.01Q3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Qwen2.5-VL-72B-InstructQ3_K_S73.4B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ3_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Chronos-Platinum-72BQ3_K_S72.7B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
UI-TARS-72B-DPOQ3_K_S73.4B32.12 GiB10.63 GiB43.77 GiB0.87 GiB13±26.5%
Step-3.5-Flash-REAP-121B-A11BI1-IQ2_XXS121B29.52 GiB13.30 GiB43.75 GiB0.89 GiB13±26.5%
GLM-4.5-Air-REAP-82B-A12BMoEIQ3_XS81.9B36.66 GiB6.11 GiB43.70 GiB0.94 GiB28±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-IQ2_M109B36.39 GiB6.38 GiB43.70 GiB0.94 GiB29±37%
Mixtral-8x22B-v0.1MoEIQ2_XXS141B35.28 GiB7.44 GiB43.68 GiB0.96 GiB18±37%
c4ai-command-r-plus-08-2024Q2_K_S104B34.08 GiB8.50 GiB43.66 GiB0.98 GiB13±26.5%
Qwen3-72B-SynthesisQ3_K_S72.7B31.95 GiB10.63 GiB43.61 GiB1.03 GiB13±26.5%
Meta-Llama-3-70B-InstructQ3_K_M70.6B31.92 GiB10.63 GiB43.57 GiB1.07 GiB13±26.5%
Maenad-70BI1-Q3_K_M70.6B31.91 GiB10.63 GiB43.56 GiB1.08 GiB13±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 W7900 run?
2024 of 2118 indexed open-weight models fit a Radeon Pro W7900 at 65,536 context with q8_0 KV cache, the largest being INTELLECT-1-Instruct at F32. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon Pro W7900 actually have?
Its nameplate is 48 GB, but about 44.64 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon Pro W7900 fast for local AI?
Its memory bandwidth is 864 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.