AMD · workstation

Radeon Pro W6800

Radeon Pro W6800 has 32 GB of VRAM at 512 GB/s — about 29.76 GiB usable after driver and compositor overhead. 2012 of 2118 indexed models fit at 8K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
32 GB
GDDR6
Bandwidth
512 GB/s
256-bit bus
Tensor FP16
—
dense
TDP
250 W
$2249 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1727vision language 181video 16image 2embedding 26audio tts 21audio asr 39

What fits at 8K context

largest quantization that fits, per model · 2012 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-IQ3_XXS79.7B28.68 GiB0.19 GiB29.76 GiB0.00 GiB70±37%
MiniCPM-o-2_6F328.7B28.37 GiB0.44 GiB29.76 GiB0.00 GiB11±26.5%
IQuest-Coder-V1-40B-InstructI1-Q5_K_M39.8B26.26 GiB2.50 GiB29.76 GiB0.00 GiB11±26.5%
GLM-4.5-Air-REAP-82B-A12BMoEIQ1_M81.9B27.38 GiB1.44 GiB29.74 GiB0.02 GiB35±37%
c4ai-command-r-plus-08-2024IQ2_XXS104B26.65 GiB2.00 GiB29.73 GiB0.03 GiB11±26.5%
NousCoder-14BBF1614.8B27.51 GiB1.25 GiB29.72 GiB0.04 GiB11±26.5%
Qwen3-14B-Claude-4.5-Opus-High-Reasoning-DistillBF1614.8B27.51 GiB1.25 GiB29.72 GiB0.04 GiB11±26.5%
Qwen3-14BBF1614.8B27.51 GiB1.25 GiB29.72 GiB0.04 GiB11±26.5%
Qwen3-14B-abliteratedBF1614.8B27.51 GiB1.25 GiB29.72 GiB0.04 GiB11±26.5%
Josiefied-Qwen3-14B-abliterated-v3BF1614.8B27.51 GiB1.25 GiB29.72 GiB0.04 GiB11±26.5%
Hermes-4-14BBF1614.8B27.51 GiB1.25 GiB29.72 GiB0.04 GiB11±26.5%
Qwen3-14B-BaseF1614.8B27.51 GiB1.25 GiB29.72 GiB0.04 GiB11±26.5%
Salience-1.5-ProMoEQ6_K_L36.0B28.66 GiB0.16 GiB29.72 GiB0.04 GiB63±37%
Qwable-v1MoEQ6_K_L36.0B28.66 GiB0.16 GiB29.72 GiB0.04 GiB63±37%
T-SearchMoEQ6_K_L36.0B28.66 GiB0.16 GiB29.72 GiB0.04 GiB63±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-TQ1_0109B27.25 GiB1.50 GiB29.68 GiB0.08 GiB38±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-IQ4_XS57.3B27.68 GiB1.05 GiB29.67 GiB0.09 GiB41±37%
Phi-3.5-MoE-instructMoEKV unresolvedQ5_K_L41.9B27.75 GiB1.00 GiB29.66 GiB0.10 GiB31±37%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.66 GiB0.10 GiB11±26.5%
Yi-34B-200K-DARE-megamerge-v8Q6_K34.4B26.78 GiB1.88 GiB29.64 GiB0.12 GiB11±26.5%
Nous-Hermes-2-Yi-34BI1-Q6_K34.4B26.78 GiB1.88 GiB29.64 GiB0.12 GiB11±26.5%
Nous-Capybara-limarpv3-34BI1-Q6_K34.4B26.78 GiB1.88 GiB29.64 GiB0.12 GiB11±26.5%
Apertus-70B-Instruct-2509UD-IQ3_XXS70.6B26.01 GiB2.50 GiB29.59 GiB0.17 GiB12±26.5%
Rombo-LLM-V3.0-Qwen-72bI1-IQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
Qwen2.5-72B-Instruct-abliteratedI1-IQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
HuatuoGPT-o1-72BIQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
MiroThinker-v1.0-72BI1-IQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
Malaysian-Qwen2.5-72B-InstructI1-IQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
Qwen2.5-72BI1-IQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
magnum-v4-72bI1-IQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
KAT-Dev-72B-ExpIQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
Homer-v1.0-Qwen2.5-72BIQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ2_S72.7B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
Qwen2.5-VL-72B-InstructIQ2_S73.4B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
UI-TARS-72B-DPOIQ2_S73.4B26.02 GiB2.50 GiB29.55 GiB0.21 GiB12±26.5%
archangel_sft-kto_llama30bIQ4_XS32.5B16.28 GiB12.19 GiB29.43 GiB0.33 GiB12±26.5%
Qwen3.5-88BMoEI1-Q2_K_S87.7B28.30 GiB0.19 GiB29.41 GiB0.35 GiB57±37%
Kimi-Linear-48B-A3B-InstructMoEQ4_K_L49.1B28.26 GiB0.24 GiB29.40 GiB0.36 GiB12±26.5%
Gemma-4-Novelist-Eclipse-31BQ6_K32.7B25.97 GiB2.42 GiB29.37 GiB0.39 GiB12±26.5%
Gemma-4-31B-StyleTuneQ6_K32.7B25.97 GiB2.42 GiB29.37 GiB0.39 GiB12±26.5%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ1_S139B26.55 GiB1.94 GiB29.37 GiB0.39 GiB37±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ1_S139B26.55 GiB1.94 GiB29.37 GiB0.39 GiB37±37%
Apriel-1.6-15b-ThinkerBF1614.9B26.88 GiB1.50 GiB29.33 GiB0.43 GiB12±26.5%
Assistant_Pepe_70BQ2_K70.6B25.79 GiB2.50 GiB29.32 GiB0.44 GiB12±26.5%
Wizard-Vicuna-30B-UncensoredI1-IQ4_XS32.5B16.15 GiB12.19 GiB29.31 GiB0.45 GiB12±26.5%
Hunyuan-A13B-InstructMoEQ2_K80.4B27.40 GiB1.00 GiB29.29 GiB0.47 GiB12±26.5%
ALIA-40b-fc-2606I1-Q5_K_M40.4B26.78 GiB1.50 GiB29.29 GiB0.47 GiB12±26.5%
ALIA-40b-instruct-2606I1-Q5_K_M40.4B26.78 GiB1.50 GiB29.29 GiB0.47 GiB12±26.5%
Qwen3-Coder-NextMoEUD-IQ3_S79.7B27.65 GiB0.75 GiB29.29 GiB0.47 GiB59±37%
Hermes-4-70BUD-IQ3_XXS70.6B25.76 GiB2.50 GiB29.28 GiB0.48 GiB12±26.5%
Llama-3.3-70B-InstructUD-IQ3_XXS70.6B25.76 GiB2.50 GiB29.28 GiB0.48 GiB12±26.5%
DeepSeek-R1-Distill-Llama-70BUD-IQ3_XXS70.6B25.76 GiB2.50 GiB29.28 GiB0.48 GiB12±26.5%
Darwin-35B-A3B-OpusMoEQ6_K_L36.0B28.22 GiB0.16 GiB29.28 GiB0.48 GiB64±37%
Aurora-Code-1MoEQ6_K_L34.7B28.22 GiB0.16 GiB29.28 GiB0.48 GiB64±37%
grug-35b-v2MoEQ6_K_L35.1B28.22 GiB0.16 GiB29.28 GiB0.48 GiB64±37%
grug-35bMoEQ6_K_L35.1B28.22 GiB0.16 GiB29.28 GiB0.48 GiB64±37%
WorldSim-Opus-3.6-35B-A3BMoEQ6_K_L35.1B28.22 GiB0.16 GiB29.28 GiB0.48 GiB64±37%
Qwen3.6-35B-A3B-AnkoMoEQ6_K_L35.1B28.22 GiB0.16 GiB29.28 GiB0.48 GiB64±37%
KAT-Coder-V2.5-DevMoEQ6_K_L34.7B28.22 GiB0.16 GiB29.28 GiB0.48 GiB64±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.

Measured on this card

third-party benchmarks, aggregated
Workload◍MedianMiddle 50%Runs
Image generation8.00 it/s7.45–8.7412
Benchmarked· n=12

Aggregated from community-submitted runs, so the spread is wide by nature — it covers different models, resolutions, step counts and settings, not one controlled configuration. Read the middle 50% rather than the median alone. These figures are reproduced with attribution from vladmandic-sd-data-benchmark, which publishes no licence — so we display and link rather than redistribute them.

Questions people ask

What AI models can a Radeon Pro W6800 run?
2012 of 2118 indexed open-weight models fit a Radeon Pro W6800 at 8,192 context with f16 KV cache, the largest being Huihui-Qwen3-Coder-Next-abliterated at I1-IQ3_XXS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon Pro W6800 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 W6800 fast for local AI?
Its memory bandwidth is 512 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.