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 32K 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 32K context

largest quantization that fits, per model · 2115 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-Coder-REAP-363B-A35BMoEQ5_K_S363B232.76 GiB4.12 GiB237.81 GiB0.27 GiB62±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ4_1402B233.50 GiB3.19 GiB237.61 GiB0.47 GiB100±37%
GLM-4.7MoEQ5_K_S358B230.45 GiB6.11 GiB237.50 GiB0.58 GiB65±37%
GLM-4.5MoEQ5_K_S358B230.45 GiB6.11 GiB237.50 GiB0.58 GiB65±37%
Hermes-3-Llama-3.1-405BQ4_K_L406B227.83 GiB8.37 GiB237.38 GiB0.70 GiB16±26.5%
MiniMax-M3MoEQ4_K_S427B234.10 GiB1.99 GiB237.00 GiB1.08 GiB85±37%
Qwen3-VL-235B-A22B-InstructMoEQ8_0236B232.77 GiB3.12 GiB236.83 GiB1.25 GiB65±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ8_0236B232.77 GiB3.12 GiB236.83 GiB1.25 GiB65±37%
Qwen3-235B-A22BMoEQ8_0235B232.77 GiB3.12 GiB236.83 GiB1.25 GiB65±37%
Qwen3-235B-A22B-Thinking-2507MoEQ8_0235B232.77 GiB3.12 GiB236.83 GiB1.25 GiB65±37%
Qwen3-235B-A22B-Instruct-2507MoEQ8_0235B232.77 GiB3.12 GiB236.83 GiB1.25 GiB65±37%
GLM-4.6-Derestricted-v3MoEQ5_K_S357B229.46 GiB6.11 GiB236.51 GiB1.57 GiB66±37%
GLM-4.6MoEQ5_K_S357B229.46 GiB6.11 GiB236.51 GiB1.57 GiB66±37%
Hermes-4-405BQ4_K_M406B226.38 GiB8.37 GiB235.93 GiB2.15 GiB16±26.5%
DeepSeek-Coder-V2-Instruct-0724MoEQ8_0236B233.41 GiB1.12 GiB235.47 GiB2.61 GiB84±37%
DeepSeek-V2.5MoEQ8_0236B233.41 GiB1.12 GiB235.47 GiB2.61 GiB84±37%
DeepSeek-Coder-V2-InstructMoEQ8_0236B233.41 GiB1.12 GiB235.47 GiB2.61 GiB84±37%
Trinity-Large-ThinkingMoEQ4_1399B232.94 GiB1.42 GiB235.28 GiB2.80 GiB101±37%
Trinity-Large-PreviewMoEQ4_1399B232.76 GiB1.42 GiB235.10 GiB2.98 GiB101±37%
Trinity-Large-TrueBaseMoEQ4_1399B232.76 GiB1.42 GiB235.10 GiB2.98 GiB101±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 GiB0.40 GiB233.56 GiB4.52 GiB91±37%
Qwen3.5-122B-A10BMoEBF16125B232.24 GiB0.40 GiB233.56 GiB4.52 GiB91±37%
ERNIE-4.5-300B-A47B-PTQ6_K300B228.88 GiB3.59 GiB233.49 GiB4.59 GiB16±26.5%
Ornith-1.0-397BMoEQ4_1397B231.99 GiB0.50 GiB233.44 GiB4.64 GiB101±37%
Nex-N2-ProMoEQ4_1397B231.99 GiB0.50 GiB233.44 GiB4.64 GiB101±37%
Qwen3.5-397B-A17BMoEQ4_K_L403B231.91 GiB0.50 GiB233.36 GiB4.72 GiB101±37%
Kimi-K2-ThinkingMoEUD-TQ1_01058B229.90 GiB1.14 GiB232.01 GiB6.07 GiB98±37%
MiniMax-M2.1MoEQ8_0229B226.44 GiB4.12 GiB231.44 GiB6.64 GiB80±37%
MiniMax-M2.5MoEQ8_0229B226.44 GiB4.12 GiB231.44 GiB6.64 GiB80±37%
MiniMax-M2MoEQ8_0229B226.44 GiB4.12 GiB231.44 GiB6.64 GiB80±37%
MiniMax-M2.7MoEQ8_0229B226.44 GiB4.12 GiB231.44 GiB6.64 GiB80±37%
DeepSeek-V3.1MoEQ2_K_L685B229.03 GiB1.14 GiB231.14 GiB6.94 GiB89±37%
DeepSeek-V3.1-TerminusMoEQ2_K_L685B229.03 GiB1.14 GiB231.14 GiB6.94 GiB89±37%
cogito-671b-v2.1MoEQ2_K_L671B229.02 GiB1.14 GiB231.14 GiB6.94 GiB89±37%
DeepSeek-V3.2MoEQ2_K_L685B228.72 GiB1.14 GiB230.83 GiB7.25 GiB89±37%
DeepSeek-R1-0528MoEQ2_K_L685B228.17 GiB1.14 GiB230.29 GiB7.79 GiB90±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ2_K_L685B228.17 GiB1.14 GiB230.29 GiB7.79 GiB90±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ2_K_L671B228.17 GiB1.14 GiB230.29 GiB7.79 GiB90±37%
DeepSeek-V3-0324MoEQ2_K_L685B228.11 GiB1.14 GiB230.23 GiB7.85 GiB90±37%
r1-1776MoEQ2_K_L671B228.11 GiB1.14 GiB230.23 GiB7.85 GiB90±37%
DeepSeek-R1MoEQ2_K_L685B228.11 GiB1.14 GiB230.23 GiB7.85 GiB90±37%
DeepSeek-Prover-V2-671BMoEQ2_K_L685B227.64 GiB1.14 GiB229.76 GiB8.32 GiB90±37%
Laguna-M.1MoEQ8_0226B223.63 GiB4.65 GiB229.22 GiB8.86 GiB65±37%
Kimi-K2-InstructMoEUD-TQ1_01026B226.87 GiB1.14 GiB228.99 GiB9.09 GiB99±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEBF16124B224.93 GiB1.46 GiB227.29 GiB10.79 GiB81±37%
MiMo-V2.5-ProMoEKV unresolvedIQ1_M1023B220.44 GiB5.81 GiB227.22 GiB10.86 GiB84±37%
GLM-5MoEUD-IQ2_XXS754B224.52 GiB1.46 GiB226.92 GiB11.16 GiB90±37%
Kimi-K2.5MoEUD-TQ1_01059B223.09 GiB1.14 GiB225.21 GiB12.87 GiB100±37%
GLM-5.2MoEUD-IQ2_M753B222.19 GiB1.46 GiB224.60 GiB13.48 GiB90±37%
GLM-4.7-REAP-218B-A32BMoEQ8_0218B216.23 GiB6.11 GiB223.28 GiB14.80 GiB57±37%
Mistral-Small-4-119B-2603MoEBF16119B221.65 GiB0.37 GiB222.95 GiB15.13 GiB95±37%
GLM-5.1MoEUD-IQ2_M754B219.96 GiB1.46 GiB222.37 GiB15.71 GiB91±37%
Laguna-S-2.1MoEBF16118B219.05 GiB0.87 GiB220.84 GiB17.24 GiB88±37%
Kimi-K2.6MoEIQ1_M1059B216.46 GiB1.14 GiB218.58 GiB19.50 GiB103±37%
Qwen3-Coder-480B-A35B-InstructMoEQ3_K_M480B213.50 GiB4.12 GiB218.55 GiB19.53 GiB74±37%
command-a-plus-05-2026-bf16MoEQ8_0219B216.03 GiB0.76 GiB217.69 GiB20.39 GiB75±37%
MiMo-V2.5MoEKV unresolvedUD-Q5_K_M311B214.37 GiB1.99 GiB217.31 GiB20.77 GiB91±37%
InklingMoEIQ1_M952B210.69 GiB4.38 GiB215.99 GiB22.09 GiB89±37%
GLM-4.6-REAP-268B-A32BMoEQ6_K269B205.68 GiB6.11 GiB212.73 GiB25.35 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.

Questions people ask

What AI models can a Instinct MI325X run?
2115 of 2118 indexed open-weight models fit a Instinct MI325X at 32,768 context with q8_0 KV cache, the largest being Qwen3-Coder-REAP-363B-A35B at Q5_K_S. 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.
Instinct MI325X — what AI models can it run locally? — ossmodeldb