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. 2039 of 2118 indexed models fit at 4K 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 1750vision language 185image 2video 16audio tts 21embedding 26audio asr 39

What fits at 4K context

largest quantization that fits, per model · 2039 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Hunyuan-A13B-InstructMoEQ4_080.4B43.44 GiB0.27 GiB44.60 GiB0.04 GiB13±26.5%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_K_S79.7B43.55 GiB0.05 GiB44.49 GiB0.15 GiB81±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q6_K49.9B38.11 GiB5.31 GiB44.46 GiB0.18 GiB13±26.5%
Valkyrie-49B-v2.1I1-Q6_K49.9B38.11 GiB5.31 GiB44.46 GiB0.18 GiB13±26.5%
Llama-3_3-Nemotron-Super-49B-v1Q6_K49.9B38.11 GiB5.31 GiB44.46 GiB0.18 GiB13±26.5%
c4ai-command-r-plus-08-2024IQ3_S104B42.80 GiB0.53 GiB44.41 GiB0.23 GiB13±26.5%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ1_S124B43.19 GiB0.18 GiB44.27 GiB0.37 GiB64±37%
Rombo-LLM-V3.0-Qwen-72bI1-Q4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
Qwen2.5-72B-Instruct-abliteratedI1-Q4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
Qwen2.5-72B-Instruct-abliterated-v2I1-Q4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
HuatuoGPT-o1-72BQ4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
MiroThinker-v1.0-72BI1-Q4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
Malaysian-Qwen2.5-72B-InstructI1-Q4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
Qwen2.5-72BI1-Q4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
Kimi-Dev-72BQ4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
KAT-Dev-72B-ExpQ4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
Homer-v1.0-Qwen2.5-72BQ4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
Qwen2.5-VL-72B-InstructQ4_173.4B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q4_172.7B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
UI-TARS-72B-DPOQ4_173.4B42.56 GiB0.66 GiB44.25 GiB0.39 GiB13±26.5%
GLM-4.5-Air-DerestrictedMoEQ2_K110B42.94 GiB0.38 GiB44.25 GiB0.39 GiB54±37%
GLM-4.5-AirMoEQ2_K110B42.94 GiB0.38 GiB44.25 GiB0.39 GiB54±37%
Behemoth-X-123B-v2Q2_K_L123B42.46 GiB0.73 GiB44.24 GiB0.40 GiB13±26.5%
Mistral-Large-Instruct-2411Q2_K_L123B42.46 GiB0.73 GiB44.24 GiB0.40 GiB13±26.5%
Qwen3.5-122B-A10BMoEQ2_K125B43.21 GiB0.05 GiB44.18 GiB0.46 GiB73±37%
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16MoEQ4_K_S75.4B43.15 GiB0.00 GiB44.12 GiB0.52 GiB102±37%
Qwen3.6-35B-A3B-REAM-160-ru-agentMoEBF1623.6B43.09 GiB0.04 GiB44.04 GiB0.60 GiB63±37%
Qwen3-Coder-NextMoEQ4_079.7B42.93 GiB0.20 GiB44.02 GiB0.62 GiB79±37%
Qwen3-Next-80B-A3B-ThinkingMoEQ4_081.3B42.93 GiB0.20 GiB44.02 GiB0.62 GiB79±37%
Qwen3-Next-80B-A3B-InstructMoEQ4_081.3B42.93 GiB0.20 GiB44.02 GiB0.62 GiB79±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ2_M139B42.58 GiB0.51 GiB43.98 GiB0.66 GiB57±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ2_M139B42.58 GiB0.51 GiB43.98 GiB0.66 GiB57±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-IQ3_XXS109B42.59 GiB0.40 GiB43.91 GiB0.73 GiB54±37%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEBF1623.0B42.85 GiB0.11 GiB43.87 GiB0.77 GiB49±37%
GLM-4.7-Flash-REAP-23B-A3BMoEBF1623.0B42.85 GiB0.11 GiB43.87 GiB0.77 GiB49±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-Q6_K57.3B42.64 GiB0.28 GiB43.86 GiB0.78 GiB55±37%
Qwen2.5-Coder-32B-InstructQ5_032.8B42.17 GiB0.53 GiB43.70 GiB0.94 GiB13±26.5%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q2_K125B42.67 GiB0.05 GiB43.64 GiB1.00 GiB74±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ2_K123B42.66 GiB0.05 GiB43.64 GiB1.00 GiB74±37%
EuroLLM-22B-Instruct-2512BF1622.6B42.17 GiB0.45 GiB43.58 GiB1.06 GiB13±26.5%
GLM-4.6VMoEQ2_K_L108B42.21 GiB0.38 GiB43.52 GiB1.12 GiB55±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q2_K121B41.19 GiB1.34 GiB43.46 GiB1.18 GiB13±26.5%
Assistant_Pepe_70BQ4_170.6B41.76 GiB0.66 GiB43.45 GiB1.19 GiB13±26.5%
Qwen3.5-88BMoEI1-Q3_K_L87.7B42.43 GiB0.05 GiB43.41 GiB1.23 GiB66±37%
Mistral-Small-Instruct-2409Q3_K_M22.2B41.93 GiB0.46 GiB43.36 GiB1.28 GiB13±26.5%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ4_K_S—42.37 GiB0.05 GiB43.31 GiB1.33 GiB83±37%
Apertus-70B-Instruct-2509Q4_K_L70.6B41.46 GiB0.66 GiB43.21 GiB1.43 GiB13±26.5%
Huihui-GLM-4.5-Air-abliterated-lossytensorsMoEI1-Q2_K110B41.88 GiB0.38 GiB43.19 GiB1.45 GiB55±37%
Qwen3-42B-A3B-2507-Thinking-Abliterated-uncensored-TOTAL-RECALL-v2-Medium-MASTER-CODERMoEQ8_042.4B41.98 GiB0.28 GiB43.15 GiB1.49 GiB56±37%
CalmeRys-78B-Orpo-v0.1I1-Q4_078.0B41.30 GiB0.71 GiB43.04 GiB1.60 GiB13±26.5%
Meta-Llama-3-70B-InstructQ4_170.6B41.28 GiB0.66 GiB42.97 GiB1.67 GiB13±26.5%
Maenad-70BI1-Q4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 GiB13±26.5%
L3.3-Electra-R1-70bI1-Q4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 GiB13±26.5%
L3.3-70B-Magnum-v4-SEQ4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 GiB13±26.5%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 GiB13±26.5%
Llama-3.3_70_b_uncensored_continuedI1-Q4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 GiB13±26.5%
grok-oss-Revenant-70BI1-Q4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 GiB13±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 GiB13±26.5%
Hermes-4-70B-hereticI1-Q4_170.6B41.27 GiB0.66 GiB42.96 GiB1.68 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?
2039 of 2118 indexed open-weight models fit a Radeon Pro W7900 at 4,096 context with q8_0 KV cache, the largest being Hunyuan-A13B-Instruct at Q4_0. 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.