AMD · workstation

Radeon Pro W7800

Radeon Pro W7800 has 32 GB of VRAM at 576 GB/s — about 29.76 GiB usable after driver and compositor overhead. 2022 of 2118 indexed models fit at 8K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
32 GB
GDDR6
Bandwidth
576 GB/s
256-bit bus
Tensor FP16
—
dense
TDP
260 W
$2499 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1737vision language 181video 16image 2embedding 26audio tts 21audio asr 39

What fits at 8K context

largest quantization that fits, per model · 2022 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Hypernova-60B-2605MoEI1-IQ3_S58.7B28.73 GiB0.15 GiB29.76 GiB0.00 GiB60±37%
Llama-3_1-Nemotron-51B-InstructQ2_K51.5B18.09 GiB10.63 GiB29.75 GiB0.01 GiB13±26.5%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q4_K_S53.0B28.13 GiB0.70 GiB29.72 GiB0.04 GiB45±37%
Seed-OSS-36B-Instruct-biprojected-norm-preserving-abliteratedI1-Q6_K36.2B27.63 GiB1.06 GiB29.69 GiB0.07 GiB13±26.5%
Seed-OSS-36B-InstructQ6_K36.2B27.63 GiB1.06 GiB29.69 GiB0.07 GiB13±26.5%
Hermes-4.3-36B-hereticI1-Q6_K36.2B27.63 GiB1.06 GiB29.69 GiB0.07 GiB13±26.5%
Hermes-4.3-36BQ6_K36.2B27.63 GiB1.06 GiB29.69 GiB0.07 GiB13±26.5%
Seed-OSS-36B-BaseQ6_K36.2B27.63 GiB1.06 GiB29.69 GiB0.07 GiB13±26.5%
Rombo-LLM-V3.0-Qwen-72bI1-IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Qwen2.5-72B-Instruct-abliteratedI1-IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
HuatuoGPT-o1-72BIQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
MiroThinker-v1.0-72BI1-IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
EVA-Qwen2.5-72B-v0.2IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Qwen2.5-Math-72B-InstructIQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Qwen2.5-72B-InstructIQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Malaysian-Qwen2.5-72B-InstructI1-IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Qwen2.5-72BI1-IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
magnum-v4-72bI1-IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
KAT-Dev-72B-ExpIQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Homer-v1.0-Qwen2.5-72BIQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Qwen2.5-VL-72B-InstructIQ2_M73.4B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Chronos-Platinum-72BIQ2_M72.7B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
UI-TARS-72B-DPOIQ2_M73.4B27.32 GiB1.33 GiB29.68 GiB0.08 GiB13±26.5%
Snowpiercer-15B-v4BF1615.0B27.90 GiB0.83 GiB29.68 GiB0.08 GiB13±26.5%
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-IQ3_XXS79.7B28.68 GiB0.10 GiB29.67 GiB0.09 GiB80±37%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.66 GiB0.10 GiB13±26.5%
Maenad-70BI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Rombos-LLM-70b-Llama-3.3I1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
L3.3-Electra-R1-70bI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Latxa-Llama-3.1-70B-Instruct-v2I1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Llama-3.3_70_b_uncensored_continuedI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Llama-3.3-70B-Instruct-abliteratedI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
grok-oss-Revenant-70BI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
L3.3-70B-Euryale-v2.3I1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Hermes-3-Llama-3.1-70BIQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Hermes-4-70B-hereticI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Llama-3.3-70B-InstructIQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Llama-3.1-70BIQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Anubis-70B-v1.2IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Hermes-4-70BIQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Golem-70B-v1bI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
DeepSeek-R1-Distill-Llama-70B-hereticI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
llama-3-firefunction-v2IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Legion-V2.1-LLaMa-70BI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Assistant_Pepe_70BI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Tess-R1-Limerick-Llama-3.1-70BIQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
New-Dawn-Llama-3-70B-32K-v1.0I1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-IQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Athene-70BIQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
L3.3-70B-Magnum-DiamondIQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Meta-Llama-3-70B-InstructIQ3_XS70.6B27.29 GiB1.33 GiB29.65 GiB0.11 GiB13±26.5%
Salience-1.5-ProMoEQ6_K_L36.0B28.66 GiB0.08 GiB29.65 GiB0.11 GiB72±37%
Qwable-v1MoEQ6_K_L36.0B28.66 GiB0.08 GiB29.65 GiB0.11 GiB72±37%
T-SearchMoEQ6_K_L36.0B28.66 GiB0.08 GiB29.65 GiB0.11 GiB72±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 W7800 run?
2022 of 2118 indexed open-weight models fit a Radeon Pro W7800 at 8,192 context with q8_0 KV cache, the largest being Hypernova-60B-2605 at I1-IQ3_S. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon Pro W7800 actually have?
Its nameplate is 32 GB, but about 29.76 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon Pro W7800 fast for local AI?
Its memory bandwidth is 576 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.
Radeon Pro W7800 — what AI models can it run locally? — ossmodeldb