AMD · datacenter

Instinct MI300X

Instinct MI300X has 192 GB of VRAM at 5300 GB/s — about 178.56 GiB usable after driver and compositor overhead. 2104 of 2118 indexed models fit at 128K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
192 GB
HBM3
Bandwidth
5300 GB/s
8192-bit bus
Tensor FP16
1307 TF
dense
TDP
750 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1808vision language 192image 2audio tts 21audio asr 39video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2104 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.6-Derestricted-v3MoEQ3_K_M357B153.01 GiB24.44 GiB178.38 GiB0.18 GiB45±37%
GLM-4.6MoEIQ3_M357B152.98 GiB24.44 GiB178.36 GiB0.20 GiB45±37%
MiMo-V2.5MoEKV unresolvedQ4_K_S311B169.41 GiB7.97 GiB178.33 GiB0.23 GiB76±37%
DeepSeek-R1-0528MoEUD-IQ1_S685B172.75 GiB4.56 GiB178.28 GiB0.28 GiB86±37%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB1.99 GiB178.23 GiB0.33 GiB105±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB4.56 GiB177.92 GiB0.64 GiB86±37%
step-3.5-flashQ6_K199B150.81 GiB26.05 GiB177.79 GiB0.77 GiB19±26.5%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB4.56 GiB177.64 GiB0.92 GiB86±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_S310B168.15 GiB7.97 GiB177.07 GiB1.49 GiB76±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB4.56 GiB177.01 GiB1.55 GiB86±37%
GLM-4.6-REAP-268B-A32BMoEQ4_K_M269B151.37 GiB24.44 GiB176.74 GiB1.82 GiB43±37%
MiniMax-M3MoEIQ3_XXS427B167.65 GiB7.97 GiB176.53 GiB2.03 GiB76±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB5.83 GiB176.10 GiB2.46 GiB82±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.05 GiB2.51 GiB96±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB1.99 GiB175.90 GiB2.66 GiB97±37%
Hermes-3-Llama-3.1-405BQ2_K_L406B140.98 GiB33.47 GiB175.63 GiB2.93 GiB19±26.5%
grok-2MoEQ4_1270B157.54 GiB17.00 GiB175.58 GiB2.98 GiB30±37%
Trinity-Large-PreviewMoEUD-IQ3_XXS399B170.01 GiB4.40 GiB175.34 GiB3.22 GiB98±37%
MiniMax-M2.7MoEUD-Q5_K_M229B157.23 GiB16.47 GiB174.59 GiB3.97 GiB59±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB12.75 GiB174.47 GiB4.09 GiB74±37%
Trinity-Large-ThinkingMoEQ3_K_M399B168.90 GiB4.40 GiB174.23 GiB4.33 GiB99±37%
Trinity-Large-TrueBaseMoEQ3_K_M399B168.74 GiB4.40 GiB174.07 GiB4.49 GiB99±37%
Hermes-4-405BQ2_K406B139.07 GiB33.47 GiB173.72 GiB4.84 GiB19±26.5%
ERNIE-4.5-300B-A47B-PTQ4_K_S300B157.93 GiB14.34 GiB173.30 GiB5.26 GiB19±26.5%
HarmonicHarlequin_v5-20BQ8_033.3B32.97 GiB138.13 GiB172.04 GiB6.52 GiB20±26.5%
Ornith-1.0-397BMoEQ3_K_M397B169.03 GiB1.99 GiB171.97 GiB6.59 GiB108±37%
Nex-N2-ProMoEQ3_K_M397B169.03 GiB1.99 GiB171.97 GiB6.59 GiB108±37%
Hy3MoEIQ4_XS299B149.73 GiB21.25 GiB171.92 GiB6.64 GiB50±37%
GLM-4.7MoEQ3_K_S358B146.52 GiB24.44 GiB171.89 GiB6.67 GiB46±37%
GLM-4.5MoEQ3_K_S358B146.52 GiB24.44 GiB171.89 GiB6.67 GiB46±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB5.83 GiB170.82 GiB7.74 GiB84±37%
GLM-5.1MoEIQ1_M754B163.93 GiB5.83 GiB170.71 GiB7.85 GiB84±37%
GLM-4.7-REAP-218B-A32BMoEQ5_K_M218B145.23 GiB24.44 GiB170.61 GiB7.95 GiB41±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB12.48 GiB168.84 GiB9.72 GiB58±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB12.48 GiB168.84 GiB9.72 GiB58±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB12.48 GiB168.77 GiB9.79 GiB58±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB12.48 GiB168.77 GiB9.79 GiB58±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB12.48 GiB168.77 GiB9.79 GiB58±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB12.48 GiB168.77 GiB9.79 GiB58±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB65.88 GiB168.65 GiB9.91 GiB25±37%
MiniMax-M2.1MoEQ5_K_M229B151.23 GiB16.47 GiB168.58 GiB9.98 GiB60±37%
MiniMax-M2MoEQ5_K_M229B151.23 GiB16.47 GiB168.58 GiB9.98 GiB60±37%
MiniMax-M2.5MoEQ5_K_M229B151.16 GiB16.47 GiB168.51 GiB10.05 GiB60±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ5_K_M229B151.16 GiB16.47 GiB168.51 GiB10.05 GiB60±37%
DeepSeek-V3.1-TerminusMoEIQ2_XXS685B162.59 GiB4.56 GiB168.13 GiB10.43 GiB89±37%
cogito-671b-v2.1MoEIQ2_XXS671B162.59 GiB4.56 GiB168.13 GiB10.43 GiB89±37%
DeepSeek-V3-0324MoEIQ2_XXS685B162.45 GiB4.56 GiB167.98 GiB10.58 GiB89±37%
r1-1776MoEIQ2_XXS671B162.45 GiB4.56 GiB167.98 GiB10.58 GiB89±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB4.56 GiB167.98 GiB10.58 GiB89±37%
DeepSeek-V3.1MoEUD-TQ1_0685B158.79 GiB4.56 GiB164.33 GiB14.23 GiB91±37%
Step-3.7-FlashUD-Q5_K_M201B136.43 GiB26.05 GiB163.41 GiB15.15 GiB21±26.5%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_S363B145.81 GiB16.47 GiB163.21 GiB15.35 GiB53±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB4.48 GiB161.16 GiB17.40 GiB88±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB4.48 GiB161.16 GiB17.40 GiB88±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB4.48 GiB161.16 GiB17.40 GiB88±37%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB8.50 GiB159.16 GiB19.40 GiB21±26.5%
Qwen2.5-72BF1672.7B135.44 GiB21.25 GiB157.72 GiB20.84 GiB21±26.5%
Kimi-Dev-72BBF1672.7B135.44 GiB21.25 GiB157.72 GiB20.84 GiB21±26.5%
Qwen2.5-VL-72B-InstructBF1673.4B135.44 GiB21.25 GiB157.72 GiB20.84 GiB21±26.5%
Qwen3-Coder-480B-A35B-InstructMoEUD-IQ1_M480B139.43 GiB16.47 GiB156.84 GiB21.72 GiB57±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.

Measured on this card

third-party benchmarks, aggregated
Workload◍MedianMiddle 50%Runs
Prompt processing11021.13 tok/s4938.37–11679.3611
Text generation169.73 tok/s159.80–225.9011
Benchmarked· n=11

Aggregated from community-submitted runs, so the spread is wide by nature — it covers different models, resolutions, step counts and settings, not one controlled configuration. Read the middle 50% rather than the median alone. These figures are reproduced with attribution from llama.cpp-discussion-14640.

Questions people ask

What AI models can a Instinct MI300X run?
2104 of 2118 indexed open-weight models fit a Instinct MI300X at 131,072 context with q8_0 KV cache, the largest being GLM-4.6-Derestricted-v3 at Q3_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Instinct MI300X actually have?
Its nameplate is 192 GB, but about 178.56 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Instinct MI300X fast for local AI?
Its memory bandwidth is 5300 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.