AMD · datacenter

Instinct MI325X

Instinct MI325X has 256 GB of VRAM at 6000 GB/s — about 238.08 GiB usable after driver and compositor overhead. 2115 of 2118 indexed models fit at 64K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
256 GB
HBM3e
Bandwidth
6000 GB/s
8192-bit bus
Tensor FP16
1307 TF
dense
TDP
1000 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1817vision language 194image 2audio tts 21audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2115 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
ERNIE-4.5-300B-A47B-PTQ6_K300B228.88 GiB7.17 GiB237.08 GiB1.00 GiB16±26.5%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB2.24 GiB236.59 GiB1.49 GiB80±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB2.24 GiB236.59 GiB1.49 GiB80±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB2.24 GiB236.59 GiB1.49 GiB80±37%
Trinity-Large-ThinkingMoEQ4_1399B232.94 GiB2.41 GiB236.28 GiB1.80 GiB96±37%
Trinity-Large-PreviewMoEQ4_1399B232.76 GiB2.41 GiB236.10 GiB1.98 GiB96±37%
Trinity-Large-TrueBaseMoEQ4_1399B232.76 GiB2.41 GiB236.10 GiB1.98 GiB96±37%
MiniMax-M3MoEUD-Q4_K_S427B230.71 GiB3.98 GiB235.61 GiB2.47 GiB80±37%
MiniMax-M2.1MoEQ8_0229B226.44 GiB8.23 GiB235.56 GiB2.52 GiB70±37%
MiniMax-M2.5MoEQ8_0229B226.44 GiB8.23 GiB235.56 GiB2.52 GiB70±37%
MiniMax-M2MoEQ8_0229B226.44 GiB8.23 GiB235.56 GiB2.52 GiB70±37%
MiniMax-M2.7MoEQ8_0229B226.44 GiB8.23 GiB235.56 GiB2.52 GiB70±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ3_S561B233.30 GiB0.00 GiB234.23 GiB3.85 GiB83±37%
Qwen3.5-122B-A10B-hereticMoEBF16123B232.24 GiB0.80 GiB233.96 GiB4.12 GiB90±37%
Qwen3.5-122B-A10BMoEBF16125B232.24 GiB0.80 GiB233.96 GiB4.12 GiB90±37%
Ornith-1.0-397BMoEQ4_1397B231.99 GiB1.00 GiB233.94 GiB4.14 GiB99±37%
Nex-N2-ProMoEQ4_1397B231.99 GiB1.00 GiB233.94 GiB4.14 GiB99±37%
Laguna-M.1MoEQ8_0226B223.63 GiB9.30 GiB233.87 GiB4.21 GiB57±37%
Qwen3.5-397B-A17BMoEQ4_K_L403B231.91 GiB1.00 GiB233.86 GiB4.22 GiB99±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ4_K_M402B226.09 GiB6.38 GiB233.40 GiB4.68 GiB89±37%
Kimi-K2-ThinkingMoEUD-TQ1_01058B229.90 GiB2.28 GiB233.15 GiB4.93 GiB93±37%
MiMo-V2.5-ProMoEKV unresolvedIQ1_M1023B220.44 GiB11.62 GiB233.03 GiB5.05 GiB69±37%
Hermes-3-Llama-3.1-405BQ4_K_S406B214.67 GiB16.73 GiB232.59 GiB5.49 GiB16±26.5%
DeepSeek-V3.1MoEQ2_K_L685B229.03 GiB2.28 GiB232.28 GiB5.80 GiB85±37%
DeepSeek-V3.1-TerminusMoEQ2_K_L685B229.03 GiB2.28 GiB232.28 GiB5.80 GiB85±37%
cogito-671b-v2.1MoEQ2_K_L671B229.02 GiB2.28 GiB232.28 GiB5.80 GiB85±37%
DeepSeek-V3.2MoEQ2_K_L685B228.72 GiB2.28 GiB231.97 GiB6.11 GiB86±37%
DeepSeek-R1-0528MoEQ2_K_L685B228.17 GiB2.28 GiB231.43 GiB6.65 GiB86±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ2_K_L685B228.17 GiB2.28 GiB231.43 GiB6.65 GiB86±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ2_K_L671B228.17 GiB2.28 GiB231.42 GiB6.66 GiB86±37%
DeepSeek-V3-0324MoEQ2_K_L685B228.11 GiB2.28 GiB231.37 GiB6.71 GiB86±37%
r1-1776MoEQ2_K_L671B228.11 GiB2.28 GiB231.37 GiB6.71 GiB86±37%
DeepSeek-R1MoEQ2_K_L685B228.11 GiB2.28 GiB231.37 GiB6.71 GiB86±37%
DeepSeek-Prover-V2-671BMoEQ2_K_L685B227.64 GiB2.28 GiB230.90 GiB7.18 GiB86±37%
Kimi-K2-InstructMoEUD-TQ1_01026B226.87 GiB2.28 GiB230.13 GiB7.95 GiB94±37%
GLM-4.7-REAP-218B-A32BMoEQ8_0218B216.23 GiB12.22 GiB229.39 GiB8.69 GiB49±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEBF16124B224.93 GiB2.92 GiB228.75 GiB9.33 GiB77±37%
GLM-5MoEUD-IQ2_XXS754B224.52 GiB2.91 GiB228.38 GiB9.70 GiB85±37%
Kimi-K2.5MoEUD-TQ1_01059B223.09 GiB2.28 GiB226.35 GiB11.73 GiB95±37%
GLM-5.2MoEUD-IQ2_M753B222.19 GiB2.91 GiB226.05 GiB12.03 GiB85±37%
GLM-5.1MoEUD-IQ2_M754B219.96 GiB2.91 GiB223.83 GiB14.25 GiB86±37%
Mistral-Small-4-119B-2603MoEBF16119B221.65 GiB0.75 GiB223.33 GiB14.75 GiB93±37%
GLM-4.7MoEQ4_1358B209.72 GiB12.22 GiB222.88 GiB15.20 GiB59±37%
GLM-4.5MoEQ4_1358B209.66 GiB12.22 GiB222.82 GiB15.26 GiB59±37%
Qwen3-Coder-480B-A35B-InstructMoEQ3_K_M480B213.50 GiB8.23 GiB222.67 GiB15.41 GiB65±37%
GLM-4.6-Derestricted-v3MoEQ4_1357B208.76 GiB12.22 GiB221.92 GiB16.16 GiB59±37%
GLM-4.6MoEQ4_1357B208.76 GiB12.22 GiB221.92 GiB16.16 GiB59±37%
Laguna-S-2.1MoEBF16118B219.05 GiB1.67 GiB221.64 GiB16.44 GiB85±37%
Qwen3-Coder-REAP-363B-A35BMoEQ4_1363B211.68 GiB8.23 GiB220.85 GiB17.23 GiB59±37%
InklingMoEIQ1_M952B210.69 GiB8.77 GiB220.37 GiB17.71 GiB76±37%
Kimi-K2.6MoEIQ1_M1059B216.46 GiB2.28 GiB219.72 GiB18.36 GiB98±37%
MiMo-V2.5MoEKV unresolvedUD-Q5_K_M311B214.37 GiB3.98 GiB219.30 GiB18.78 GiB84±37%
GLM-4.6-REAP-268B-A32BMoEQ6_K269B205.68 GiB12.22 GiB218.84 GiB19.24 GiB55±37%
command-a-plus-05-2026-bf16MoEQ8_0219B216.03 GiB1.29 GiB218.22 GiB19.86 GiB74±37%
grok-2MoEQ6_K270B206.17 GiB8.50 GiB215.71 GiB22.37 GiB31±37%
GLM-4.5-AirMoEBF16110B205.82 GiB6.11 GiB212.86 GiB25.22 GiB66±37%
Step-3.7-FlashQ8_0201B197.44 GiB13.30 GiB211.66 GiB26.42 GiB18±26.5%
Hy3MoEQ5_K_M299B198.22 GiB10.63 GiB209.78 GiB28.30 GiB66±37%
step-3.5-flashQ8_0199B195.04 GiB13.30 GiB209.26 GiB28.82 GiB18±26.5%
MiMo-V2-FlashMoEKV unresolvedQ5_K_M310B204.17 GiB3.98 GiB209.10 GiB28.98 GiB87±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 Instinct MI325X run?
2115 of 2118 indexed open-weight models fit a Instinct MI325X at 65,536 context with q8_0 KV cache, the largest being ERNIE-4.5-300B-A47B-PT at Q6_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a Instinct MI325X actually have?
Its nameplate is 256 GB, but about 238.08 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Instinct MI325X fast for local AI?
Its memory bandwidth is 6000 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.