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. 2038 of 2118 indexed models fit at 32K 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
$3499 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1749vision language 185image 2video 16audio tts 21embedding 26audio asr 39

What fits at 32K context

largest quantization that fits, per model · 2038 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-72B-SynthesisQ4_K_S72.7B40.80 GiB2.81 GiB44.64 GiB0.00 GiB13±26.5%
Llama-3_1-Nemotron-51B-InstructQ3_K_S51.5B21.10 GiB22.50 GiB44.64 GiB0.00 GiB13±26.5%
Step-3.7-FlashIQ1_S201B40.03 GiB3.67 GiB44.63 GiB0.01 GiB13±26.5%
Apertus-70B-Instruct-2509Q4_K_M70.6B40.72 GiB2.81 GiB44.61 GiB0.03 GiB13±26.5%
GLM-4.5-Air-DerestrictedMoEIQ2_M110B42.02 GiB1.62 GiB44.56 GiB0.08 GiB45±37%
GLM-4.5-AirMoEIQ2_M110B42.02 GiB1.62 GiB44.56 GiB0.08 GiB45±37%
Mistral-Medium-3.5-128BIQ2_XS128B40.41 GiB3.09 GiB44.56 GiB0.08 GiB13±26.5%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedIQ3_XXS109B41.87 GiB1.69 GiB44.48 GiB0.16 GiB45±37%
Huihui-GLM-4.5-Air-abliterated-lossytensorsMoEI1-Q2_K110B41.88 GiB1.62 GiB44.43 GiB0.21 GiB45±37%
Trinity-2-Codestral-22B-v0.2F1622.2B41.44 GiB1.97 GiB44.37 GiB0.27 GiB13±26.5%
Mistral-Small-Drummer-22BF1622.2B41.44 GiB1.97 GiB44.37 GiB0.27 GiB13±26.5%
Cydonia-v1.3-Magnum-v4-22BF1622.2B41.44 GiB1.97 GiB44.37 GiB0.27 GiB13±26.5%
Mistral-Small-Instruct-2409F1622.2B41.44 GiB1.97 GiB44.37 GiB0.27 GiB13±26.5%
Mistral-Small-22B-ArliAI-RPMax-v1.1F1622.2B41.44 GiB1.97 GiB44.37 GiB0.27 GiB13±26.5%
magnum-v4-22bF1622.2B41.44 GiB1.97 GiB44.37 GiB0.27 GiB13±26.5%
Codestral-22B-v0.1BF1622.2B41.44 GiB1.97 GiB44.37 GiB0.27 GiB13±26.5%
dolphin-2.9.1-mixtral-1x22bMoEBF1622.2B41.42 GiB1.97 GiB44.35 GiB0.29 GiB7±37%
Qwen3.5-122B-A10BMoEQ2_K125B43.21 GiB0.21 GiB44.34 GiB0.30 GiB71±37%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEBF1623.0B42.85 GiB0.46 GiB44.23 GiB0.41 GiB47±37%
GLM-4.7-Flash-REAP-23B-A3BMoEBF1623.0B42.85 GiB0.46 GiB44.23 GiB0.41 GiB47±37%
XORTRON-NXTXPRTXXLI1-Q2_K_S128B40.05 GiB3.09 GiB44.20 GiB0.44 GiB13±26.5%
GLM-4.6VMoEQ2_K108B41.64 GiB1.62 GiB44.19 GiB0.45 GiB46±37%
Devstral-2-123B-Instruct-2512IQ2_M125B40.03 GiB3.09 GiB44.18 GiB0.46 GiB13±26.5%
Qwen3.6-35B-A3B-REAM-160-ru-agentMoEBF1623.6B43.09 GiB0.18 GiB44.17 GiB0.47 GiB61±37%
L3.3-Electra-R1-70bQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Llama-3.3-70B-Instruct-abliteratedQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Llama-3.3-70B-InstructQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
L3.3-70B-Euryale-v2.3Q4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Rombos-LLM-70b-Llama-3.3Q4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Anubis-70B-v1.2Q4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Tess-R1-Limerick-Llama-3.1-70BQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
functionary-medium-v3.2KV unresolvedQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Athene-70BQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Infinity-Instruct-7M-Gen-Llama3_1-70BQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
Hermes-3-Llama-3.1-70BQ4_K_L70.6B40.33 GiB2.81 GiB44.16 GiB0.48 GiB13±26.5%
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16MoEQ4_K_S75.4B43.15 GiB0.00 GiB44.12 GiB0.52 GiB102±37%
Qwen3-42B-A3B-2507-Thinking-Abliterated-uncensored-TOTAL-RECALL-v2-Medium-MASTER-CODERMoEQ8_042.4B41.98 GiB1.18 GiB44.05 GiB0.59 GiB48±37%
Midnight-Miqu-70B-v1.5I1-Q4_169.0B40.20 GiB2.81 GiB44.04 GiB0.60 GiB13±26.5%
Llama-3_3-Nemotron-Super-49B-v1_5Q3_K_S49.9B20.45 GiB22.50 GiB43.99 GiB0.65 GiB13±26.5%
Valkyrie-49B-v2.1I1-IQ3_S49.9B20.45 GiB22.50 GiB43.99 GiB0.65 GiB13±26.5%
Llama-3_3-Nemotron-Super-49B-v1Q3_K_S49.9B20.45 GiB22.50 GiB43.99 GiB0.65 GiB13±26.5%
c4ai-command-r-plus-08-2024IQ3_XS104B40.61 GiB2.25 GiB43.94 GiB0.70 GiB13±26.5%
Qwen3-Coder-NextMoEIQ4_NL79.7B42.20 GiB0.84 GiB43.93 GiB0.71 GiB69±37%
Qwen3-Next-80B-A3B-ThinkingMoEIQ4_NL81.3B42.20 GiB0.84 GiB43.93 GiB0.71 GiB69±37%
Qwen3-Next-80B-A3B-InstructMoEIQ4_NL81.3B42.20 GiB0.84 GiB43.93 GiB0.71 GiB69±37%
L3-DARKEST-PLANET-16.5BQ6_K16.5B40.47 GiB2.50 GiB43.91 GiB0.73 GiB13±26.5%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_079.7B42.78 GiB0.21 GiB43.89 GiB0.75 GiB79±37%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-IQ3_XS109B41.25 GiB1.69 GiB43.87 GiB0.77 GiB45±37%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q2_K125B42.67 GiB0.21 GiB43.80 GiB0.84 GiB71±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ2_K123B42.66 GiB0.21 GiB43.80 GiB0.84 GiB71±37%
CalmeRys-78B-Orpo-v0.1I1-IQ4_XS78.0B39.63 GiB3.02 GiB43.69 GiB0.95 GiB13±26.5%
calme-2.3-rys-78bIQ4_XS78.0B39.63 GiB3.02 GiB43.69 GiB0.95 GiB13±26.5%
Qwen3.5-88BMoEI1-Q3_K_L87.7B42.43 GiB0.21 GiB43.57 GiB1.07 GiB64±37%
Phi-3.5-MoE-instructMoEKV unresolvedQ8_041.9B41.44 GiB1.13 GiB43.48 GiB1.16 GiB36±37%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ4_K_S—42.37 GiB0.21 GiB43.47 GiB1.17 GiB80±37%
Meta-Llama-3-70B-InstructQ4_K_M70.6B39.61 GiB2.81 GiB43.45 GiB1.19 GiB13±26.5%
Maenad-70BI1-Q4_K_M70.6B39.60 GiB2.81 GiB43.44 GiB1.20 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q4_K_M70.6B39.60 GiB2.81 GiB43.44 GiB1.20 GiB13±26.5%
calme-2.4-llama3-70bQ4_K_M70.6B39.60 GiB2.81 GiB43.44 GiB1.20 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 Dual Slot run?
2038 of 2118 indexed open-weight models fit a Radeon Pro W7900 Dual Slot at 32,768 context with q4_0 KV cache, the largest being Qwen3-72B-Synthesis 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.