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 8K 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 8K context

largest quantization that fits, per model · 2039 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.5-Air-DerestrictedMoEQ2_K110B42.94 GiB0.76 GiB44.63 GiB0.01 GiB51±37%
GLM-4.5-AirMoEQ2_K110B42.94 GiB0.76 GiB44.63 GiB0.01 GiB51±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q5_K_M49.9B32.96 GiB10.63 GiB44.63 GiB0.01 GiB13±26.5%
Valkyrie-49B-v2.1I1-Q5_K_M49.9B32.96 GiB10.63 GiB44.63 GiB0.01 GiB13±26.5%
Llama-3_3-Nemotron-Super-49B-v1Q5_K_M49.9B32.96 GiB10.63 GiB44.63 GiB0.01 GiB13±26.5%
Behemoth-X-123B-v2Q2_K123B42.09 GiB1.46 GiB44.61 GiB0.03 GiB13±26.5%
Mistral-Large-Instruct-2411Q2_K123B42.09 GiB1.46 GiB44.61 GiB0.03 GiB13±26.5%
Huihui-Qwen3-Coder-Next-abliteratedMoEQ4_K_S79.7B43.55 GiB0.10 GiB44.54 GiB0.10 GiB80±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ2_M139B42.58 GiB1.03 GiB44.49 GiB0.15 GiB52±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ2_M139B42.58 GiB1.03 GiB44.49 GiB0.15 GiB52±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEIQ1_S124B43.19 GiB0.37 GiB44.46 GiB0.18 GiB62±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-IQ3_XXS109B42.59 GiB0.80 GiB44.31 GiB0.33 GiB51±37%
Step-3.5-Flash-REAP-121B-A11BI1-Q2_K121B41.19 GiB2.14 GiB44.26 GiB0.38 GiB13±26.5%
Qwen3.5-122B-A10BMoEQ2_K125B43.21 GiB0.10 GiB44.23 GiB0.41 GiB72±37%
Qwen2.5-Coder-32B-InstructQ5_032.8B42.17 GiB1.06 GiB44.23 GiB0.41 GiB13±26.5%
Qwen3-Coder-NextMoEQ4_079.7B42.93 GiB0.40 GiB44.22 GiB0.42 GiB76±37%
Qwen3-Next-80B-A3B-ThinkingMoEQ4_081.3B42.93 GiB0.40 GiB44.22 GiB0.42 GiB76±37%
Qwen3-Next-80B-A3B-InstructMoEQ4_081.3B42.93 GiB0.40 GiB44.22 GiB0.42 GiB76±37%
Hunyuan-A13B-InstructMoEIQ4_NL80.4B42.77 GiB0.53 GiB44.19 GiB0.45 GiB13±26.5%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-Q6_K57.3B42.64 GiB0.56 GiB44.14 GiB0.50 GiB53±37%
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16MoEQ4_K_S75.4B43.15 GiB0.00 GiB44.12 GiB0.52 GiB102±37%
Assistant_Pepe_70BQ4_170.6B41.76 GiB1.33 GiB44.11 GiB0.53 GiB13±26.5%
Qwen3.6-35B-A3B-REAM-160-ru-agentMoEBF1623.6B43.09 GiB0.08 GiB44.08 GiB0.56 GiB62±37%
EuroLLM-22B-Instruct-2512BF1622.6B42.17 GiB0.90 GiB44.03 GiB0.61 GiB13±26.5%
GLM-4.7-Flash-REAP-23B-A3B-absolute-heresyMoEBF1623.0B42.85 GiB0.22 GiB43.98 GiB0.66 GiB48±37%
GLM-4.7-Flash-REAP-23B-A3BMoEBF1623.0B42.85 GiB0.22 GiB43.98 GiB0.66 GiB48±37%
GLM-4.6VMoEQ2_K_L108B42.21 GiB0.76 GiB43.90 GiB0.74 GiB51±37%
Apertus-70B-Instruct-2509Q4_K_L70.6B41.46 GiB1.33 GiB43.87 GiB0.77 GiB13±26.5%
Mistral-Small-Instruct-2409Q3_K_M22.2B41.93 GiB0.93 GiB43.82 GiB0.82 GiB13±26.5%
CalmeRys-78B-Orpo-v0.1I1-Q4_078.0B41.30 GiB1.43 GiB43.75 GiB0.89 GiB13±26.5%
OYM-Qimi-122B-A10B-K2.6MoEI1-Q2_K125B42.67 GiB0.10 GiB43.69 GiB0.95 GiB73±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoEQ2_K123B42.66 GiB0.10 GiB43.69 GiB0.95 GiB73±37%
Meta-Llama-3-70B-InstructQ4_170.6B41.28 GiB1.33 GiB43.63 GiB1.01 GiB13±26.5%
Maenad-70BI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
L3.3-Electra-R1-70bI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
L3.3-70B-Magnum-v4-SEQ4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Llama-3.3_70_b_uncensored_continuedI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
grok-oss-Revenant-70BI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Hermes-4-70B-hereticI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Hermes-4-70BQ4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Llama-3.3-70B-Instruct-abliteratedQ4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Llama-3.1-70BQ4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Llama-3.3-70B-InstructQ4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Anubis-70B-v1.2Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Golem-70B-v1bI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70BQ4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Legion-V2.1-LLaMa-70BI1-Q4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
SEMIKONG-70BQ4_170.6B41.27 GiB1.33 GiB43.62 GiB1.02 GiB13±26.5%
Mistral-Medium-3.5-128BUD-IQ2_M128B41.08 GiB1.46 GiB43.60 GiB1.04 GiB13±26.5%
Huihui-GLM-4.5-Air-abliterated-lossytensorsMoEI1-Q2_K110B41.88 GiB0.76 GiB43.58 GiB1.06 GiB52±37%
Qwen3.5-88BMoEI1-Q3_K_L87.7B42.43 GiB0.10 GiB43.46 GiB1.18 GiB66±37%
Qwen3-42B-A3B-2507-Thinking-Abliterated-uncensored-TOTAL-RECALL-v2-Medium-MASTER-CODERMoEQ8_042.4B41.98 GiB0.56 GiB43.43 GiB1.21 GiB53±37%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ4_K_S—42.37 GiB0.10 GiB43.36 GiB1.28 GiB82±37%
Trinity-2-Codestral-22B-v0.2F1622.2B41.44 GiB0.93 GiB43.33 GiB1.31 GiB13±26.5%
Mistral-Small-Drummer-22BF1622.2B41.44 GiB0.93 GiB43.33 GiB1.31 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 8,192 context with q8_0 KV cache, the largest being GLM-4.5-Air-Derestricted at Q2_K. 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.