AMD · workstation

Radeon Pro W7900 Dual Slot

Radeon Pro W7900 Dual Slot has 48 GB of VRAM at 864 GB/s — about 44.64 GiB usable after driver and compositor overhead. 2037 of 2118 indexed models fit at 16K 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
$3499 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1749vision language 184image 2video 16audio tts 21embedding 26audio asr 39

What fits at 16K context

largest quantization that fits, per model · 2037 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_K_S79.7B43.55 GiB0.20 GiB44.64 GiB0.00 GiB78±37%
Assistant_Pepe_70BQ4_K_M70.6B40.95 GiB2.66 GiB44.63 GiB0.01 GiB13±26.5%
Qwen3-Coder-NextMoEQ4_079.7B42.93 GiB0.80 GiB44.62 GiB0.02 GiB69±37%
Qwen3-Next-80B-A3B-ThinkingMoEQ4_081.3B42.93 GiB0.80 GiB44.62 GiB0.02 GiB69±37%
Qwen3-Next-80B-A3B-InstructMoEQ4_081.3B42.93 GiB0.80 GiB44.62 GiB0.02 GiB69±37%
Rombo-LLM-V3.0-Qwen-72bI1-Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Qwen2.5-72B-Instruct-abliteratedI1-Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Qwen2.5-72B-Instruct-abliterated-v2I1-Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
HuatuoGPT-o1-72BQ4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
MiroThinker-v1.0-72BI1-Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
EVA-Qwen2.5-72B-v0.2Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Malaysian-Qwen2.5-72B-InstructI1-Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Qwen2.5-72BI1-Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
magnum-v4-72bI1-Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
KAT-Dev-72B-ExpQ4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Homer-v1.0-Qwen2.5-72BQ4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Chuluun-Qwen2.5-72B-v0.01Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Qwen2.5-VL-72B-InstructQ4_K_S73.4B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q4_K_S72.7B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
UI-TARS-72B-DPOQ4_K_S73.4B40.88 GiB2.66 GiB44.56 GiB0.08 GiB13±26.5%
Devstral-2-123B-Instruct-2512UD-IQ2_M125B40.55 GiB2.92 GiB44.53 GiB0.11 GiB13±26.5%
Qwen3-72B-SynthesisQ4_K_S72.7B40.80 GiB2.66 GiB44.48 GiB0.16 GiB13±26.5%
GLM-4.5-Air-DerestrictedMoEIQ2_M110B42.02 GiB1.53 GiB44.48 GiB0.16 GiB46±37%
GLM-4.5-AirMoEIQ2_M110B42.02 GiB1.53 GiB44.47 GiB0.17 GiB46±37%
Apertus-70B-Instruct-2509Q4_K_M70.6B40.72 GiB2.66 GiB44.46 GiB0.18 GiB13±26.5%
Mistral-Medium-3.5-128BIQ2_XS128B40.41 GiB2.92 GiB44.39 GiB0.25 GiB13±26.5%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ3_XXS109B41.87 GiB1.59 GiB44.39 GiB0.25 GiB46±37%
Huihui-GLM-4.5-Air-abliterated-lossytensorsMoEI1-Q2_K110B41.88 GiB1.53 GiB44.34 GiB0.30 GiB46±37%
Qwen3.5-122B-A10BMoEQ2_K125B43.21 GiB0.20 GiB44.33 GiB0.31 GiB71±37%
Trinity-2-Codestral-22B-v0.2F1622.2B41.44 GiB1.86 GiB44.26 GiB0.38 GiB13±26.5%
Mistral-Small-Drummer-22BF1622.2B41.44 GiB1.86 GiB44.26 GiB0.38 GiB13±26.5%
Cydonia-v1.3-Magnum-v4-22BF1622.2B41.44 GiB1.86 GiB44.26 GiB0.38 GiB13±26.5%
Mistral-Small-Instruct-2409F1622.2B41.44 GiB1.86 GiB44.26 GiB0.38 GiB13±26.5%
Mistral-Small-22B-ArliAI-RPMax-v1.1F1622.2B41.44 GiB1.86 GiB44.26 GiB0.38 GiB13±26.5%
magnum-v4-22bF1622.2B41.44 GiB1.86 GiB44.26 GiB0.38 GiB13±26.5%
Codestral-22B-v0.1BF1622.2B41.44 GiB1.86 GiB44.26 GiB0.38 GiB13±26.5%
dolphin-2.9.1-mixtral-1x22bMoEBF1622.2B41.42 GiB1.86 GiB44.24 GiB0.40 GiB7±37%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEBF1623.0B42.85 GiB0.44 GiB44.20 GiB0.44 GiB47±37%
GLM-4.7-Flash-REAP-23B-A3BMoEBF1623.0B42.85 GiB0.44 GiB44.20 GiB0.44 GiB47±37%
Llama-3_1-Nemotron-51B-InstructIQ3_M51.5B21.88 GiB21.25 GiB44.17 GiB0.47 GiB13±26.5%
Qwen3.6-35B-A3B-REAM-160-ru-agentMoEBF1623.6B43.09 GiB0.17 GiB44.16 GiB0.48 GiB61±37%
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16MoEQ4_K_S75.4B43.15 GiB0.00 GiB44.12 GiB0.52 GiB102±37%
GLM-4.6VMoEQ2_K108B41.64 GiB1.53 GiB44.10 GiB0.54 GiB46±37%
XORTRON-NXTXPRTXXLI1-Q2_K_S128B40.05 GiB2.92 GiB44.03 GiB0.61 GiB13±26.5%
L3.3-Electra-R1-70bQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Llama-3.3-70B-Instruct-abliteratedQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Llama-3.3-70B-InstructQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
L3.3-70B-Euryale-v2.3Q4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Rombos-LLM-70b-Llama-3.3Q4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Anubis-70B-v1.2Q4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Tess-R1-Limerick-Llama-3.1-70BQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
functionary-medium-v3.2KV unresolvedQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Athene-70BQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Infinity-Instruct-7M-Gen-Llama3_1-70BQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Hermes-3-Llama-3.1-70BQ4_K_L70.6B40.33 GiB2.66 GiB44.01 GiB0.63 GiB13±26.5%
Qwen3-42B-A3B-2507-Thinking-Abliterated-uncensored-TOTAL-RECALL-v2-Medium-MASTER-CODERMoEQ8_042.4B41.98 GiB1.11 GiB43.99 GiB0.65 GiB49±37%
Midnight-Miqu-70B-v1.5I1-Q4_169.0B40.20 GiB2.66 GiB43.88 GiB0.76 GiB13±26.5%
c4ai-command-r-plus-08-2024IQ3_XS104B40.61 GiB2.13 GiB43.81 GiB0.83 GiB13±26.5%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q2_K125B42.67 GiB0.20 GiB43.79 GiB0.85 GiB71±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 Pro W7900 Dual Slot run?
2037 of 2118 indexed open-weight models fit a Radeon Pro W7900 Dual Slot at 16,384 context with q8_0 KV cache, the largest being Huihui-Qwen3-Coder-Next-abliterated at Q4_K_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon Pro W7900 Dual Slot 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 Dual Slot 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.