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. 2023 of 2118 indexed models fit at 128K context with q4_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 1735vision language 184image 2video 16audio tts 21audio asr 39embedding 26

What fits at 128K context

largest quantization that fits, per model · 2023 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
reka-flash-3BF1620.9B38.94 GiB4.64 GiB44.56 GiB0.08 GiB13±26.5%
reka-flash-3.1BF1620.9B38.94 GiB4.64 GiB44.56 GiB0.08 GiB13±26.5%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-Q2_K109B36.85 GiB6.75 GiB44.52 GiB0.12 GiB28±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_079.7B42.78 GiB0.84 GiB44.52 GiB0.12 GiB69±37%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q2_K125B42.67 GiB0.84 GiB44.44 GiB0.20 GiB62±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ2_K123B42.66 GiB0.84 GiB44.44 GiB0.20 GiB62±37%
Rombo-LLM-V3.0-Qwen-72bI1-IQ3_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Qwen2.5-72B-Instruct-abliteratedI1-IQ3_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ3_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
HuatuoGPT-o1-72BQ3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
MiroThinker-v1.0-72BI1-IQ3_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
EVA-Qwen2.5-72B-v0.2Q3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Qwen2.5-Math-72B-InstructQ3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Qwen2.5-72B-InstructQ3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Malaysian-Qwen2.5-72B-InstructI1-IQ3_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Qwen2.5-72BI1-IQ3_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
magnum-v4-72bI1-IQ3_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Kimi-Dev-72BQ3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
KAT-Dev-72B-ExpQ3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Homer-v1.0-Qwen2.5-72BQ3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Chuluun-Qwen2.5-72B-v0.01Q3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Qwen2.5-VL-72B-InstructQ3_K_S73.4B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ3_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Chronos-Platinum-72BQ3_K_S72.7B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
UI-TARS-72B-DPOQ3_K_S73.4B32.12 GiB11.25 GiB44.40 GiB0.24 GiB13±26.5%
Delphi-25B-SimpleRL-MathI1-IQ1_M25.0B5.74 GiB37.65 GiB44.37 GiB0.27 GiB13±26.5%
GLM-4.5-Air-DerestrictedMoEIQ2_XXS110B36.90 GiB6.47 GiB44.30 GiB0.34 GiB28±37%
GLM-4.5-AirMoEIQ2_XXS110B36.90 GiB6.47 GiB44.29 GiB0.35 GiB28±37%
Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingQ8_039.5B39.90 GiB3.38 GiB44.24 GiB0.40 GiB13±26.5%
Step-3.5-Flash-REAP-121B-A11BI1-IQ2_XXS121B29.52 GiB13.79 GiB44.24 GiB0.40 GiB13±26.5%
Hunyuan-A13B-InstructMoEQ3_K_L80.4B38.84 GiB4.50 GiB44.24 GiB0.40 GiB13±26.5%
Qwen3-72B-SynthesisQ3_K_S72.7B31.95 GiB11.25 GiB44.23 GiB0.41 GiB13±26.5%
Qwen3.5-88BMoEI1-Q3_K_L87.7B42.43 GiB0.84 GiB44.21 GiB0.43 GiB57±37%
Meta-Llama-3-70B-InstructQ3_K_M70.6B31.92 GiB11.25 GiB44.20 GiB0.44 GiB13±26.5%
Maenad-70BI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
calme-2.4-llama3-70bQ3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
calme-2.2-llama3-70bQ3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Rombos-LLM-70b-Llama-3.3I1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
L3.3-Electra-R1-70bI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
L3.3-70B-Magnum-v4-SEQ3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Llama-3.3_70_b_uncensored_continuedI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Llama-3.3-70B-Instruct-abliteratedI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Strawberrylemonade-L3-70B-v1.2Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
grok-oss-Revenant-70BI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
L3.3-70B-Euryale-v2.3I1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Hermes-4-70B-hereticI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Hermes-4-70BQ3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Llama-3.3-70B-InstructQ3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Llama-3.1-70BQ3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Hermes-3-Llama-3.1-70BQ3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Anubis-70B-v1.2Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Golem-70B-v1bI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70BQ3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
llama-3-firefunction-v2Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 GiB13±26.5%
Legion-V2.1-LLaMa-70BI1-Q3_K_M70.6B31.91 GiB11.25 GiB44.19 GiB0.45 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?
2023 of 2118 indexed open-weight models fit a Radeon Pro W7900 at 131,072 context with q4_0 KV cache, the largest being reka-flash-3 at BF16. 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.