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. 1857 of 2118 indexed models fit at 128K context with f16 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 1577vision language 177image 1audio asr 39audio tts 21embedding 26video 16

What fits at 128K context

largest quantization that fits, per model · 1857 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
QwQ-32BQ2_K_L32.8B11.64 GiB32.00 GiB44.64 GiB0.00 GiB13±26.5%
Qwen3.5-99BMoEI1-IQ3_M99.0B40.70 GiB3.00 GiB44.63 GiB0.01 GiB43±37%
internlm2-math-plus-20bQ8_019.9B19.66 GiB24.00 GiB44.62 GiB0.02 GiB13±26.5%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-Q3_K_L57.3B26.94 GiB16.73 GiB44.61 GiB0.03 GiB16±37%
Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-UncensoredMoEI1-Q6_K33.6B25.69 GiB18.00 GiB44.60 GiB0.04 GiB13±37%
Phi-3.5-MoE-instructMoEKV unresolvedQ5_K_M41.9B27.68 GiB16.00 GiB44.58 GiB0.06 GiB15±37%
Nous-Hermes-2-Yi-34BQ2_K34.4B13.56 GiB30.00 GiB44.54 GiB0.10 GiB13±26.5%
Capybara-Tess-Yi-34B-200KQ2_K34.4B13.56 GiB30.00 GiB44.54 GiB0.10 GiB13±26.5%
OrionStar-Yi-34B-Chat-LlamaQ2_K34.4B13.56 GiB30.00 GiB44.54 GiB0.10 GiB13±26.5%
Nous-Capybara-limarpv3-34BQ2_K34.4B13.56 GiB30.00 GiB44.54 GiB0.10 GiB13±26.5%
Qwen2.5-Coder-14B-InstructQ5_K_M14.8B19.57 GiB24.00 GiB44.52 GiB0.12 GiB13±26.5%
OpenBuddy-R1-0528-Distill-Qwen3-32B-Preview0-QATQ2_K32.8B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
Qwen3-VL-32B-Instruct-ultra-uncensored-hereticI1-Q2_K33.4B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
Huihui-Qwen3-VL-32B-Instruct-abliteratedI1-Q2_K33.4B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
KAT-DevQ2_K32.8B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
ColorGUI-32BI1-Q2_K33.4B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
Qwen3-VL-32B-InstructQ2_K33.4B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
Qwen3-VL-32B-ThinkingQ2_K33.4B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
Qwen3-32B-UncensoredI1-Q2_K32.8B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
Qwen3-32BQ2_K32.8B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
Qwen3-32B-abliteratedI1-Q2_K32.8B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
DeepSWE-PreviewQ2_K32.8B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
AReaL-boba-2-32BI1-Q2_K32.8B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
Assistant_Pepe_32BI1-Q2_K32.8B11.50 GiB32.00 GiB44.49 GiB0.15 GiB13±26.5%
CallerQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Dumpling-Qwen2.5-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
OREAL-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Baichuan-M2-32B-abliteratedQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
QwQ-32B-Preview-abliterated-linear25I1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
openhands-lm-32b-v0.1I1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-Coder-32B-abliteratedI1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
INTELLECT-2Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
LongWriter-Zero-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
m1-32bI1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
XMainframe-v2-Instruct-32bI1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-Coder-32B-Python-SpecialistI1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-32b-RP-InkI1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
OpenCodeReasoning-Nemotron-32B-IOIQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-Coder-32B-Instruct-abliteratedQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
OlympicCoder-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
OpenCodeReasoning-Nemotron-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
OpenThinker-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
QwQ-32B-ArliAI-RpR-v4Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-Coder-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
QwQ-32B-abliteratedQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
DeepSeek-R1-Distill-Qwen-32B-hereticI1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
InnoSpark-HPC-RM-32BI1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
OpenThinker2-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-32B-InstructQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-Coder-32B-Instruct-UncensoredI1-Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
QwQ-32B-PreviewQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
DeepSeek-R1-Distill-Qwen-32B-abliteratedQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
TinyR1-32B-PreviewQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
deepseek-r1-qwen-2.5-32B-ablatedQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Rombos-LLM-V2.5-Qwen-32bQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-32B-ArliAI-RPMax-v1.3Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
DeepSeek-R1-Distill-Qwen-32B-Blunt-UncensoredQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
DeepSeek-R1-Distill-Qwen-32BQ2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
Qwen2.5-VL-32B-InstructQ2_K33.5B11.47 GiB32.00 GiB44.47 GiB0.17 GiB13±26.5%
EVA-Qwen2.5-32B-v0.2Q2_K32.8B11.47 GiB32.00 GiB44.47 GiB0.17 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?
1857 of 2118 indexed open-weight models fit a Radeon Pro W7900 at 131,072 context with f16 KV cache, the largest being QwQ-32B at Q2_K_L. 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.