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 16K 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 16K context

largest quantization that fits, per model · 2115 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiMo-V2-FlashMoEKV unresolvedQ6_K310B236.11 GiB1.00 GiB238.06 GiB0.02 GiB88±37%
Devstral-2-123B-Instruct-2512BF16125B232.89 GiB2.92 GiB236.87 GiB1.21 GiB16±26.5%
Mistral-Medium-3.5-128BBF16128B232.89 GiB2.92 GiB236.87 GiB1.21 GiB16±26.5%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ4_1402B233.50 GiB1.59 GiB236.02 GiB2.06 GiB108±37%
MiniMax-M3MoEQ4_K_S427B234.10 GiB1.00 GiB236.01 GiB2.07 GiB88±37%
Qwen3-Coder-REAP-363B-A35BMoEQ5_K_S363B232.76 GiB2.06 GiB235.75 GiB2.33 GiB66±37%
Qwen3-VL-235B-A22B-InstructMoEQ8_0236B232.77 GiB1.56 GiB235.27 GiB2.81 GiB69±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ8_0236B232.77 GiB1.56 GiB235.27 GiB2.81 GiB69±37%
Qwen3-235B-A22BMoEQ8_0235B232.77 GiB1.56 GiB235.27 GiB2.81 GiB69±37%
Qwen3-235B-A22B-Thinking-2507MoEQ8_0235B232.77 GiB1.56 GiB235.27 GiB2.81 GiB69±37%
Qwen3-235B-A22B-Instruct-2507MoEQ8_0235B232.77 GiB1.56 GiB235.27 GiB2.81 GiB69±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB0.56 GiB234.91 GiB3.17 GiB85±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB0.56 GiB234.91 GiB3.17 GiB85±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB0.56 GiB234.91 GiB3.17 GiB85±37%
Trinity-Large-ThinkingMoEQ4_1399B232.94 GiB0.92 GiB234.78 GiB3.30 GiB103±37%
Trinity-Large-PreviewMoEQ4_1399B232.76 GiB0.92 GiB234.60 GiB3.48 GiB103±37%
Trinity-Large-TrueBaseMoEQ4_1399B232.76 GiB0.92 GiB234.60 GiB3.48 GiB103±37%
GLM-4.7MoEQ5_K_S358B230.45 GiB3.05 GiB234.45 GiB3.63 GiB72±37%
GLM-4.5MoEQ5_K_S358B230.45 GiB3.05 GiB234.45 GiB3.63 GiB72±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ3_S561B233.30 GiB0.00 GiB234.23 GiB3.85 GiB83±37%
GLM-4.6-Derestricted-v3MoEQ5_K_S357B229.46 GiB3.05 GiB233.45 GiB4.63 GiB72±37%
GLM-4.6MoEQ5_K_S357B229.46 GiB3.05 GiB233.45 GiB4.63 GiB72±37%
Qwen3.5-122B-A10B-hereticMoEBF16123B232.24 GiB0.20 GiB233.36 GiB4.72 GiB92±37%
Qwen3.5-122B-A10BMoEBF16125B232.24 GiB0.20 GiB233.36 GiB4.72 GiB92±37%
Hermes-3-Llama-3.1-405BQ4_K_L406B227.83 GiB4.18 GiB233.19 GiB4.89 GiB16±26.5%
Ornith-1.0-397BMoEQ4_1397B231.99 GiB0.25 GiB233.19 GiB4.89 GiB102±37%
Nex-N2-ProMoEQ4_1397B231.99 GiB0.25 GiB233.19 GiB4.89 GiB102±37%
Qwen3.5-397B-A17BMoEQ4_K_L403B231.91 GiB0.25 GiB233.11 GiB4.97 GiB102±37%
Hermes-4-405BQ4_K_M406B226.38 GiB4.18 GiB231.74 GiB6.34 GiB17±26.5%
ERNIE-4.5-300B-A47B-PTQ6_K300B228.88 GiB1.79 GiB231.70 GiB6.38 GiB17±26.5%
Kimi-K2-ThinkingMoEUD-TQ1_01058B229.90 GiB0.57 GiB231.44 GiB6.64 GiB100±37%
DeepSeek-V3.1MoEQ2_K_L685B229.03 GiB0.57 GiB230.57 GiB7.51 GiB91±37%
DeepSeek-V3.1-TerminusMoEQ2_K_L685B229.03 GiB0.57 GiB230.57 GiB7.51 GiB91±37%
cogito-671b-v2.1MoEQ2_K_L671B229.02 GiB0.57 GiB230.57 GiB7.51 GiB91±37%
DeepSeek-V3.2MoEQ2_K_L685B228.72 GiB0.57 GiB230.26 GiB7.82 GiB92±37%
DeepSeek-R1-0528MoEQ2_K_L685B228.17 GiB0.57 GiB229.72 GiB8.36 GiB92±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ2_K_L685B228.17 GiB0.57 GiB229.72 GiB8.36 GiB92±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ2_K_L671B228.17 GiB0.57 GiB229.72 GiB8.36 GiB92±37%
DeepSeek-V3-0324MoEQ2_K_L685B228.11 GiB0.57 GiB229.66 GiB8.42 GiB92±37%
r1-1776MoEQ2_K_L671B228.11 GiB0.57 GiB229.66 GiB8.42 GiB92±37%
DeepSeek-R1MoEQ2_K_L685B228.11 GiB0.57 GiB229.66 GiB8.42 GiB92±37%
MiniMax-M2.1MoEQ8_0229B226.44 GiB2.06 GiB229.38 GiB8.70 GiB87±37%
MiniMax-M2.5MoEQ8_0229B226.44 GiB2.06 GiB229.38 GiB8.70 GiB87±37%
MiniMax-M2MoEQ8_0229B226.44 GiB2.06 GiB229.38 GiB8.70 GiB87±37%
MiniMax-M2.7MoEQ8_0229B226.44 GiB2.06 GiB229.38 GiB8.70 GiB87±37%
DeepSeek-Prover-V2-671BMoEQ2_K_L685B227.64 GiB0.57 GiB229.19 GiB8.89 GiB92±37%
Kimi-K2-InstructMoEUD-TQ1_01026B226.87 GiB0.57 GiB228.42 GiB9.66 GiB101±37%
Laguna-M.1MoEQ8_0226B223.63 GiB2.32 GiB226.90 GiB11.18 GiB69±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEBF16124B224.93 GiB0.73 GiB226.56 GiB11.52 GiB84±37%
GLM-5MoEUD-IQ2_XXS754B224.52 GiB0.73 GiB226.19 GiB11.89 GiB92±37%
Kimi-K2.5MoEUD-TQ1_01059B223.09 GiB0.57 GiB224.64 GiB13.44 GiB103±37%
MiMo-V2.5-ProMoEKV unresolvedIQ1_M1023B220.44 GiB2.91 GiB224.31 GiB13.77 GiB94±37%
GLM-5.2MoEUD-IQ2_M753B222.19 GiB0.73 GiB223.87 GiB14.21 GiB93±37%
Mistral-Small-4-119B-2603MoEBF16119B221.65 GiB0.19 GiB222.77 GiB15.31 GiB96±37%
GLM-5.1MoEUD-IQ2_M754B219.96 GiB0.73 GiB221.64 GiB16.44 GiB94±37%
Laguna-S-2.1MoEBF16118B219.05 GiB0.47 GiB220.45 GiB17.63 GiB89±37%
GLM-4.7-REAP-218B-A32BMoEQ8_0218B216.23 GiB3.05 GiB220.23 GiB17.85 GiB61±37%
Kimi-K2.6MoEIQ1_M1059B216.46 GiB0.57 GiB218.01 GiB20.07 GiB105±37%
command-a-plus-05-2026-bf16MoEQ8_0219B216.03 GiB0.49 GiB217.42 GiB20.66 GiB76±37%
Qwen3-Coder-480B-A35B-InstructMoEQ3_K_M480B213.50 GiB2.06 GiB216.50 GiB21.58 GiB79±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 16,384 context with q8_0 KV cache, the largest being MiMo-V2-Flash 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.