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. 2106 of 2118 indexed models fit at 128K context with f16 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 1809vision language 193image 2audio tts 21audio asr 39embedding 26video 16

What fits at 128K context

largest quantization that fits, per model · 2106 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
DeepSeek-V3.2MoEQ2_K685B228.51 GiB8.58 GiB238.07 GiB0.01 GiB69±37%
DeepSeek-R1-0528MoEQ2_K_L685B228.17 GiB8.58 GiB237.73 GiB0.35 GiB69±37%
DeepSeek-TNG-R1T2-ChimeraMoEQ2_K_L685B228.17 GiB8.58 GiB237.73 GiB0.35 GiB69±37%
cogito-v2-preview-deepseek-671B-MoEMoEQ2_K_L671B228.17 GiB8.58 GiB237.72 GiB0.36 GiB69±37%
DeepSeek-V3-0324MoEQ2_K_L685B228.11 GiB8.58 GiB237.67 GiB0.41 GiB69±37%
r1-1776MoEQ2_K_L671B228.11 GiB8.58 GiB237.67 GiB0.41 GiB69±37%
DeepSeek-R1MoEQ2_K_L685B228.11 GiB8.58 GiB237.67 GiB0.41 GiB69±37%
DeepSeek-Prover-V2-671BMoEQ2_K_L685B227.64 GiB8.58 GiB237.20 GiB0.88 GiB69±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ4_K_S402B212.15 GiB24.00 GiB237.08 GiB1.00 GiB54±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEBF16124B224.93 GiB11.00 GiB236.83 GiB1.25 GiB60±37%
Ornith-1.0-397BMoEQ4_1397B231.99 GiB3.75 GiB236.69 GiB1.39 GiB88±37%
Nex-N2-ProMoEQ4_1397B231.99 GiB3.75 GiB236.69 GiB1.39 GiB88±37%
GLM-4.5MoEIQ4_NL358B189.74 GiB46.00 GiB236.68 GiB1.40 GiB32±37%
Qwen3.5-397B-A17BMoEQ4_K_L403B231.91 GiB3.75 GiB236.61 GiB1.47 GiB88±37%
GLM-4.7MoEIQ4_NL358B189.62 GiB46.00 GiB236.56 GiB1.52 GiB32±37%
GLM-5MoEUD-IQ2_XXS754B224.52 GiB10.97 GiB236.43 GiB1.65 GiB65±37%
Kimi-K2-InstructMoEUD-TQ1_01026B226.87 GiB8.58 GiB236.43 GiB1.65 GiB74±37%
Qwen3-Coder-REAP-363B-A35BMoEQ4_K_M363B204.25 GiB31.00 GiB236.18 GiB1.90 GiB38±37%
Qwen3.5-122B-A10B-hereticMoEBF16123B232.24 GiB3.00 GiB236.17 GiB1.91 GiB82±37%
Qwen3.5-122B-A10BMoEBF16125B232.24 GiB3.00 GiB236.16 GiB1.92 GiB82±37%
GLM-4.6-Derestricted-v3MoEIQ4_NL357B188.93 GiB46.00 GiB235.87 GiB2.21 GiB32±37%
GLM-4.6MoEIQ4_NL357B188.93 GiB46.00 GiB235.87 GiB2.21 GiB32±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ3_M480B203.44 GiB31.00 GiB235.37 GiB2.71 GiB40±37%
Trinity-Large-ThinkingMoEQ4_K_L399B225.66 GiB8.29 GiB234.88 GiB3.20 GiB78±37%
Trinity-Large-PreviewMoEQ4_K_L399B225.46 GiB8.29 GiB234.68 GiB3.40 GiB78±37%
Trinity-Large-TrueBaseMoEQ4_K_L399B225.46 GiB8.29 GiB234.68 GiB3.40 GiB78±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ3_S561B233.30 GiB0.00 GiB234.23 GiB3.85 GiB83±37%
GLM-5.2MoEUD-IQ2_M753B222.19 GiB10.97 GiB234.11 GiB3.97 GiB65±37%
Hermes-3-Llama-3.1-405BIQ3_M406B169.26 GiB63.00 GiB233.44 GiB4.64 GiB16±26.5%
Hy3MoEQ5_K_S299B191.88 GiB40.00 GiB232.82 GiB5.26 GiB36±37%
Kimi-K2.5MoEUD-TQ1_01059B223.09 GiB8.58 GiB232.64 GiB5.44 GiB75±37%
GLM-5.1MoEUD-IQ2_M754B219.96 GiB10.97 GiB231.88 GiB6.20 GiB66±37%
MiMo-V2.5MoEKV unresolvedUD-Q5_K_M311B214.37 GiB15.00 GiB230.32 GiB7.76 GiB59±37%
MiniMax-M3MoEIQ4_XS427B213.91 GiB15.00 GiB229.83 GiB8.25 GiB60±37%
GLM-4.5-AirMoEBF16110B205.82 GiB23.00 GiB229.75 GiB8.33 GiB44±37%
DeepSeek-V3.1MoEUD-IQ2_M685B219.34 GiB8.58 GiB228.89 GiB9.19 GiB71±37%
WizardLM-Uncensored-SuperCOT-StoryTelling-30bQ8_032.5B32.19 GiB195.00 GiB228.16 GiB9.92 GiB17±26.5%
archangel_sft-kto_llama30bQ8_032.5B32.19 GiB195.00 GiB228.16 GiB9.92 GiB17±26.5%
Wizard-Vicuna-30B-UncensoredQ8_032.5B32.19 GiB195.00 GiB228.16 GiB9.92 GiB17±26.5%
dots.llm1.instMoEQ5_K_M143B101.84 GiB124.00 GiB226.77 GiB11.31 GiB16±37%
Laguna-S-2.1MoEBF16118B219.05 GiB6.14 GiB226.11 GiB11.97 GiB73±37%
Kimi-K2.6MoEIQ1_M1059B216.46 GiB8.58 GiB226.02 GiB12.06 GiB76±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedBF16109B200.76 GiB24.00 GiB225.68 GiB12.40 GiB44±37%
ERNIE-4.5-300B-A47B-PTQ5_K_M300B197.41 GiB27.00 GiB225.43 GiB12.65 GiB17±26.5%
Mistral-Small-4-119B-2603MoEBF16119B221.65 GiB2.81 GiB225.39 GiB12.69 GiB86±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_M269B177.71 GiB46.00 GiB224.65 GiB13.43 GiB31±37%
GLM-4.6VMoEBF16108B199.08 GiB23.00 GiB223.01 GiB15.07 GiB45±37%
DeepSeek-V3.1-TerminusMoEUD-IQ2_M685B213.17 GiB8.58 GiB222.72 GiB15.36 GiB72±37%
cogito-671b-v2.1MoEUD-IQ2_M671B212.81 GiB8.58 GiB222.37 GiB15.71 GiB72±37%
command-a-plus-05-2026-bf16MoEQ8_0219B216.03 GiB4.42 GiB221.35 GiB16.73 GiB67±37%
MiMo-V2-FlashMoEKV unresolvedQ5_K_M310B204.17 GiB15.00 GiB220.11 GiB17.97 GiB61±37%
Solar-Open2-250BMoEQ6_K250B191.44 GiB24.00 GiB216.36 GiB21.72 GiB51±37%
MiniMax-M2.7MoEQ6_K229B183.52 GiB31.00 GiB215.41 GiB22.67 GiB44±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB46.00 GiB214.51 GiB23.57 GiB31±37%
grok-2MoEQ5_K_M270B178.41 GiB32.00 GiB211.46 GiB26.62 GiB25±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB49.03 GiB210.16 GiB27.92 GiB18±26.5%
MiniMax-M2.1MoEQ6_K229B174.91 GiB31.00 GiB206.80 GiB31.28 GiB45±37%
MiniMax-M2MoEQ6_K229B174.91 GiB31.00 GiB206.80 GiB31.28 GiB45±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB31.00 GiB206.75 GiB31.33 GiB45±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB31.00 GiB206.75 GiB31.33 GiB45±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?
2106 of 2118 indexed open-weight models fit a Instinct MI325X at 131,072 context with f16 KV cache, the largest being DeepSeek-V3.2 at Q2_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.