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. 2108 of 2118 indexed models fit at 64K 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
vision language 192text 1812image 2audio tts 21audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Ornith-1.0-397BMoEIQ3_M397B176.54 GiB1.00 GiB178.49 GiB0.07 GiB111±37%
Nex-N2-ProMoEQ3_K_L397B176.46 GiB1.00 GiB178.41 GiB0.15 GiB111±37%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB1.00 GiB177.23 GiB1.33 GiB111±37%
DeepSeek-V3.1-TerminusMoEUD-IQ1_S685B173.84 GiB2.28 GiB177.10 GiB1.46 GiB95±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB2.28 GiB176.71 GiB1.85 GiB95±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB2.28 GiB176.38 GiB2.18 GiB95±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB2.28 GiB176.37 GiB2.19 GiB95±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.05 GiB2.51 GiB96±37%
DeepSeek-R1-0528MoEUD-IQ1_S685B172.75 GiB2.28 GiB176.01 GiB2.55 GiB95±37%
ERNIE-4.5-300B-A47B-PTQ4_K_M300B167.78 GiB7.17 GiB175.98 GiB2.58 GiB19±26.5%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB3.98 GiB175.96 GiB2.60 GiB89±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB2.28 GiB175.65 GiB2.91 GiB95±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB2.28 GiB175.36 GiB3.20 GiB96±37%
dots.llm1.instMoEQ8_0143B141.37 GiB32.94 GiB175.24 GiB3.32 GiB39±37%
Hy3MoEQ4_K_S299B163.41 GiB10.63 GiB174.97 GiB3.59 GiB66±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB1.00 GiB174.91 GiB3.65 GiB103±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB2.28 GiB174.74 GiB3.82 GiB96±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB13.30 GiB174.43 GiB4.13 GiB19±26.5%
MiMo-V2.5MoEKV unresolvedQ4_K_S311B169.41 GiB3.98 GiB174.34 GiB4.22 GiB89±37%
Trinity-Large-PreviewMoEUD-IQ3_XXS399B170.01 GiB2.41 GiB173.35 GiB5.21 GiB109±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB2.91 GiB173.19 GiB5.37 GiB94±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_S310B168.15 GiB3.98 GiB173.08 GiB5.48 GiB90±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB6.38 GiB172.95 GiB5.61 GiB85±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB12.22 GiB172.67 GiB5.89 GiB61±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB12.22 GiB172.67 GiB5.89 GiB61±37%
MiniMax-M3MoEIQ3_XXS427B167.65 GiB3.98 GiB172.55 GiB6.01 GiB90±37%
Trinity-Large-ThinkingMoEQ3_K_M399B168.90 GiB2.41 GiB172.24 GiB6.32 GiB110±37%
Trinity-Large-TrueBaseMoEQ3_K_M399B168.74 GiB2.41 GiB172.08 GiB6.48 GiB110±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB12.22 GiB172.04 GiB6.52 GiB61±37%
Qwen3-Coder-480B-A35B-InstructMoEQ2_K_L480B162.87 GiB8.23 GiB172.03 GiB6.53 GiB69±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB12.22 GiB171.41 GiB7.15 GiB61±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB8.23 GiB170.74 GiB7.82 GiB64±37%
GLM-4.6-REAP-268B-A32BMoEQ4_1269B156.99 GiB12.22 GiB170.15 GiB8.41 GiB57±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB6.38 GiB168.10 GiB10.46 GiB98±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB2.91 GiB167.91 GiB10.65 GiB96±37%
GLM-5.1MoEIQ1_M754B163.93 GiB2.91 GiB167.79 GiB10.77 GiB96±37%
grok-2MoEQ4_1270B157.54 GiB8.50 GiB167.08 GiB11.48 GiB34±37%
MiniMax-M2.7MoEUD-Q5_K_M229B157.23 GiB8.23 GiB166.35 GiB12.21 GiB78±37%
r1-1776MoEIQ2_XXS671B162.45 GiB2.28 GiB165.71 GiB12.85 GiB100±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB2.28 GiB165.71 GiB12.85 GiB100±37%
step-3.5-flashQ6_K199B150.81 GiB13.30 GiB165.04 GiB13.52 GiB20±26.5%
Hermes-3-Llama-3.1-405BIQ3_XXS406B145.14 GiB16.73 GiB163.06 GiB15.50 GiB21±26.5%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB6.24 GiB162.60 GiB15.96 GiB71±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB6.24 GiB162.60 GiB15.96 GiB71±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB6.24 GiB162.53 GiB16.03 GiB71±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB6.24 GiB162.53 GiB16.03 GiB71±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB6.24 GiB162.53 GiB16.03 GiB71±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB6.24 GiB162.53 GiB16.03 GiB71±37%
DeepSeek-V3.1MoEUD-TQ1_0685B158.79 GiB2.28 GiB162.05 GiB16.51 GiB101±37%
MiniMax-M2.1MoEQ5_K_M229B151.23 GiB8.23 GiB160.35 GiB18.21 GiB80±37%
MiniMax-M2MoEQ5_K_M229B151.23 GiB8.23 GiB160.35 GiB18.21 GiB80±37%
MiniMax-M2.5MoEQ5_K_M229B151.16 GiB8.23 GiB160.28 GiB18.28 GiB80±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ5_K_M229B151.16 GiB8.23 GiB160.28 GiB18.28 GiB80±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB2.24 GiB158.92 GiB19.64 GiB98±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB2.24 GiB158.92 GiB19.64 GiB98±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB2.24 GiB158.92 GiB19.64 GiB98±37%
GLM-4.7-REAP-218B-A32BMoEQ5_K_M218B145.23 GiB12.22 GiB158.39 GiB20.17 GiB56±37%
Hermes-4-405BQ2_K406B139.07 GiB16.73 GiB156.98 GiB21.58 GiB21±26.5%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB4.25 GiB154.91 GiB23.65 GiB22±26.5%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.03 GiB154.31 GiB24.25 GiB127±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?
2108 of 2118 indexed open-weight models fit a Instinct MI300X at 65,536 context with q8_0 KV cache, the largest being Ornith-1.0-397B at IQ3_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.