AMD · workstation

Radeon Pro W7600

Radeon Pro W7600 has 8 GB of VRAM at 288 GB/s — about 7.44 GiB usable after driver and compositor overhead. 1399 of 2118 indexed models fit at 4K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
8 GB
GDDR6
Bandwidth
288 GB/s
128-bit bus
Tensor FP16
—
dense
TDP
130 W
$599 MSRP
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1204vision language 101image 2video 7embedding 26audio tts 21audio asr 38

What fits at 4K context

largest quantization that fits, per model · 1399 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Ling-liteMoEIQ2_M16.8B6.33 GiB0.22 GiB7.44 GiB0.00 GiB81±37%
Nexa-AI-4x4B-InstructMoEI1-Q3_K_L12.1B5.96 GiB0.56 GiB7.44 GiB0.00 GiB28±37%
Bonsai-8B-unpackedQ5_K_L8.2B5.94 GiB0.56 GiB7.43 GiB0.01 GiB28±26.5%
LFM2.5-Queen-Opus-4.7-8B-A1BMoEI1-Q6_K8.5B6.48 GiB0.05 GiB7.43 GiB0.01 GiB79±37%
LFM2.5-8B-A1B-KO-SFTMoEI1-Q6_K8.5B6.48 GiB0.05 GiB7.43 GiB0.01 GiB79±37%
LFM2.5-8B-A1B-hereticMoEI1-Q6_K8.5B6.48 GiB0.05 GiB7.43 GiB0.01 GiB79±37%
LFM2.5-8B-A1B-SOMPOA-heresyMoEI1-Q6_K8.5B6.48 GiB0.05 GiB7.43 GiB0.01 GiB79±37%
Huihui-LFM2.5-8B-A1B-abliteratedMoEI1-Q6_K8.5B6.48 GiB0.05 GiB7.43 GiB0.01 GiB79±37%
Supertron2.1-8B-A1BMoEI1-Q6_K8.5B6.48 GiB0.05 GiB7.43 GiB0.01 GiB79±37%
LFM2.5-8B-A1BMoEQ6_K8.5B6.48 GiB0.05 GiB7.43 GiB0.01 GiB79±37%
gemma-7bI1-Q4_K_S8.5B4.70 GiB1.75 GiB7.42 GiB0.02 GiB28±26.5%
Ministral-3-8B-Instruct-2512-BF16Q5_K_L8.9B5.95 GiB0.53 GiB7.42 GiB0.02 GiB28±26.5%
dolphin-2.9.2-Phi-3-MediumKV unresolvedIQ3_S14.0B5.68 GiB0.78 GiB7.42 GiB0.02 GiB28±26.5%
Cydonia-v1.3-Magnum-v4-22BI1-IQ2_XXS22.2B5.58 GiB0.88 GiB7.42 GiB0.02 GiB28±26.5%
Mistral-Small-22B-ArliAI-RPMax-v1.1I1-IQ2_XXS22.2B5.58 GiB0.88 GiB7.42 GiB0.02 GiB28±26.5%
magnum-v4-22bI1-IQ2_XXS22.2B5.58 GiB0.88 GiB7.42 GiB0.02 GiB28±26.5%
Mistral-Small-Instruct-2409IQ2_XXS22.2B5.58 GiB0.88 GiB7.42 GiB0.02 GiB28±26.5%
Codestral-22B-v0.1IQ2_XXS22.2B5.58 GiB0.88 GiB7.42 GiB0.02 GiB28±26.5%
Ministral-3-14B-Instruct-2512-BF16-abliteratedI1-IQ3_M13.9B5.84 GiB0.63 GiB7.42 GiB0.02 GiB28±26.5%
Ministral-3-14B-Instruct-2512-BF16IQ3_M13.9B5.84 GiB0.63 GiB7.42 GiB0.02 GiB28±26.5%
Ministral-3-14B-Reasoning-2512-UncensoredI1-IQ3_M13.9B5.84 GiB0.63 GiB7.42 GiB0.02 GiB28±26.5%
Wan2.1-T2V-1.3BQ4_01.4B6.50 GiB0.00 GiB7.42 GiB0.02 GiB28±26.5%
dolphin-2.9.1-mixtral-1x22bMoEI1-IQ2_XXS22.2B5.58 GiB0.88 GiB7.42 GiB0.02 GiB16±37%
Qwen3.6-35B-A3B-REAM-160-ru-agentMoEIQ2_XXS23.6B6.43 GiB0.08 GiB7.42 GiB0.02 GiB115±37%
codegeex4-all-9bIQ3_XXS9.4B3.97 GiB2.50 GiB7.41 GiB0.03 GiB28±26.5%
zeta-2Q5_K_L8.3B5.97 GiB0.50 GiB7.41 GiB0.03 GiB28±26.5%
glm-4-9b-chatIQ3_XXS9.4B3.97 GiB2.50 GiB7.41 GiB0.03 GiB28±26.5%
Gemma-4-12B-StyleTuneI1-IQ3_M13.0B5.74 GiB0.72 GiB7.41 GiB0.03 GiB28±26.5%
gemma-4-12b-heretic-styletune-headI1-IQ3_M12.0B5.74 GiB0.72 GiB7.41 GiB0.03 GiB28±26.5%
syrian-gemma-12bI1-IQ3_M13.0B5.74 GiB0.72 GiB7.41 GiB0.03 GiB28±26.5%
SambaLingo-Japanese-ChatI1-Q5_K_S6.9B4.48 GiB2.00 GiB7.40 GiB0.04 GiB28±26.5%
Ling-mini-2.0MoEIQ3_XS16.3B6.35 GiB0.16 GiB7.40 GiB0.04 GiB121±37%
UncensoredLM-DeepSeek-R1-Distill-Qwen-14BIQ3_XS14.2B5.73 GiB0.72 GiB7.40 GiB0.04 GiB28±26.5%
INTELLECT-1-InstructI1-Q4_K_M10.2B5.80 GiB0.66 GiB7.40 GiB0.04 GiB28±26.5%
Qwythos-9B-v2Q4_K_L9.7B6.34 GiB0.13 GiB7.40 GiB0.04 GiB28±26.5%
Tess-4-9BQ4_K_L9.7B6.34 GiB0.13 GiB7.40 GiB0.04 GiB28±26.5%
Qwen3.5-9B-BaseQ5_19.7B6.33 GiB0.13 GiB7.39 GiB0.05 GiB28±26.5%
v6-Finch-7B-HFQ4_07.6B4.45 GiB2.00 GiB7.39 GiB0.05 GiB28±26.5%
rwkv-6-world-7bQ4_07.6B4.45 GiB2.00 GiB7.39 GiB0.05 GiB28±26.5%
Teuken-7B-instruct-research-v0.4Q6_K_L7.5B6.33 GiB0.13 GiB7.39 GiB0.05 GiB28±26.5%
internlm2-math-plus-20bI1-IQ2_XS19.9B5.68 GiB0.75 GiB7.39 GiB0.05 GiB28±26.5%
LocateAnything-3BBF163.8B6.34 GiB0.14 GiB7.39 GiB0.05 GiB28±26.5%
deepseek-math-7b-instructQ5_K_M6.9B4.59 GiB1.88 GiB7.39 GiB0.05 GiB28±26.5%
deepseek-llm-7b-chatQ5_K_M6.9B4.59 GiB1.88 GiB7.39 GiB0.05 GiB28±26.5%
Janus-Pro-7BI1-Q5_K_M7.4B4.59 GiB1.88 GiB7.39 GiB0.05 GiB28±26.5%
deepseek-coder-7b-instruct-v1.5I1-Q5_K_M6.9B4.59 GiB1.88 GiB7.39 GiB0.05 GiB28±26.5%
Phi-3-medium-128k-instructQ3_K_S14.0B5.65 GiB0.78 GiB7.39 GiB0.05 GiB28±26.5%
Phi-3-medium-4k-instructI1-IQ3_S14.0B5.65 GiB0.78 GiB7.39 GiB0.05 GiB28±26.5%
Qwen2.5-3B-Instruct-abliteratedF163.1B6.33 GiB0.14 GiB7.39 GiB0.05 GiB28±26.5%
GRM-Kerlin-3bF163.4B6.33 GiB0.14 GiB7.39 GiB0.05 GiB28±26.5%
Nous-Hermes-2-SOLAR-10.7BQ4_K_S10.7B5.70 GiB0.75 GiB7.39 GiB0.05 GiB28±26.5%
SOLAR-10.7B-Instruct-v1.0I1-Q4_K_S10.7B5.70 GiB0.75 GiB7.39 GiB0.05 GiB28±26.5%
Garnet-OCR-3B-0422F164.1B6.33 GiB0.14 GiB7.38 GiB0.06 GiB28±26.5%
Rocinante-XL-16B-v1IQ2_S16.1B5.59 GiB0.84 GiB7.38 GiB0.06 GiB28±26.5%
granite-4.1-8bQ5_K_M8.8B5.82 GiB0.63 GiB7.38 GiB0.06 GiB28±26.5%
deepseek-coder-6.7b-instructQ5_K_M6.7B4.46 GiB2.00 GiB7.38 GiB0.06 GiB28±26.5%
deepseek-coder-6.7b-baseQ5_K_M6.7B4.46 GiB2.00 GiB7.38 GiB0.06 GiB28±26.5%
deepseek-coder-6.7B-kexerI1-Q5_K_M6.7B4.46 GiB2.00 GiB7.38 GiB0.06 GiB28±26.5%
Magicoder-S-DS-6.7BI1-Q5_K_M6.7B4.46 GiB2.00 GiB7.38 GiB0.06 GiB28±26.5%
MathCoder2-CodeLlama-7BQ5_K_M6.7B4.45 GiB2.00 GiB7.38 GiB0.06 GiB28±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 W7600 run?
1399 of 2118 indexed open-weight models fit a Radeon Pro W7600 at 4,096 context with f16 KV cache, the largest being Ling-lite at IQ2_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Radeon Pro W7600 actually have?
Its nameplate is 8 GB, but about 7.44 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Radeon Pro W7600 fast for local AI?
Its memory bandwidth is 288 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.