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 32K context with q4_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 1812vision language 192image 2audio tts 21audio asr 39video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiMo-V2.5MoEKV unresolvedQ4_K_L311B176.24 GiB1.05 GiB178.24 GiB0.32 GiB100±37%
MiniMax-M2.7MoEUD-Q6_K229B175.15 GiB2.18 GiB178.21 GiB0.35 GiB95±37%
DeepSeek-R1-0528MoEIQ2_S685B176.61 GiB0.60 GiB178.19 GiB0.37 GiB102±37%
DeepSeek-V3.1-TerminusMoEIQ2_S685B176.61 GiB0.60 GiB178.19 GiB0.37 GiB102±37%
DeepSeek-V3.1MoEIQ2_S685B176.61 GiB0.60 GiB178.19 GiB0.37 GiB102±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.26 GiB178.15 GiB0.41 GiB115±37%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB0.75 GiB178.11 GiB0.45 GiB117±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.40 GiB178.10 GiB0.46 GiB81±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB0.75 GiB177.98 GiB0.58 GiB117±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB2.18 GiB177.98 GiB0.58 GiB95±37%
MiniMax-M2MoEQ6_K229B174.91 GiB2.18 GiB177.98 GiB0.58 GiB95±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB2.18 GiB177.93 GiB0.63 GiB95±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB2.18 GiB177.93 GiB0.63 GiB95±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB0.75 GiB177.92 GiB0.64 GiB117±37%
Ornith-1.0-397BMoEIQ3_M397B176.54 GiB0.26 GiB177.76 GiB0.80 GiB116±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB1.90 GiB177.49 GiB1.07 GiB19±26.5%
GLM-4.6-REAP-268B-A32BMoEQ5_K_S269B172.64 GiB3.23 GiB176.82 GiB1.74 GiB72±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_L310B174.72 GiB1.05 GiB176.72 GiB1.84 GiB100±37%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB0.26 GiB176.50 GiB2.06 GiB116±37%
grok-2MoEQ5_K_S270B173.10 GiB2.25 GiB176.39 GiB2.17 GiB35±37%
MiniMax-M3MoEQ3_K_S427B174.39 GiB1.05 GiB176.36 GiB2.20 GiB101±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.05 GiB2.51 GiB96±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB0.60 GiB175.03 GiB3.53 GiB103±37%
Hermes-3-Llama-3.1-405BIQ3_M406B169.26 GiB4.43 GiB174.87 GiB3.69 GiB19±26.5%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB0.60 GiB174.70 GiB3.86 GiB103±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB0.60 GiB174.69 GiB3.87 GiB103±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.26 GiB174.17 GiB4.39 GiB107±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB0.60 GiB173.97 GiB4.59 GiB104±37%
Hy3MoEQ4_K_L299B169.99 GiB2.81 GiB173.74 GiB4.82 GiB87±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB0.60 GiB173.68 GiB4.88 GiB104±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB0.60 GiB173.06 GiB5.50 GiB104±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB1.05 GiB173.03 GiB5.53 GiB102±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB3.23 GiB171.74 GiB6.82 GiB67±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB0.77 GiB171.05 GiB7.51 GiB104±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB1.69 GiB168.26 GiB10.30 GiB107±37%
Qwen3-Coder-480B-A35B-InstructMoEQ2_K_L480B162.87 GiB2.18 GiB165.98 GiB12.58 GiB87±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB0.77 GiB165.77 GiB12.79 GiB107±37%
GLM-5.1MoEIQ1_M754B163.93 GiB0.77 GiB165.65 GiB12.91 GiB107±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB3.67 GiB164.79 GiB13.77 GiB20±26.5%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB2.18 GiB164.69 GiB13.87 GiB79±37%
r1-1776MoEIQ2_XXS671B162.45 GiB0.60 GiB164.03 GiB14.53 GiB109±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB0.60 GiB164.03 GiB14.53 GiB109±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB3.23 GiB163.68 GiB14.88 GiB85±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB3.23 GiB163.68 GiB14.88 GiB85±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB1.69 GiB163.41 GiB15.15 GiB127±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB3.23 GiB163.06 GiB15.50 GiB85±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB3.23 GiB162.42 GiB16.14 GiB85±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB1.65 GiB158.01 GiB20.55 GiB85±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB1.65 GiB158.01 GiB20.55 GiB85±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB1.65 GiB157.94 GiB20.62 GiB85±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB1.65 GiB157.94 GiB20.62 GiB85±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB1.65 GiB157.94 GiB20.62 GiB85±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB1.65 GiB157.94 GiB20.62 GiB85±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB0.59 GiB157.27 GiB21.29 GiB107±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB0.59 GiB157.27 GiB21.29 GiB107±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB0.59 GiB157.27 GiB21.29 GiB107±37%
step-3.5-flashQ6_K199B150.81 GiB3.67 GiB155.41 GiB23.15 GiB22±26.5%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.02 GiB154.29 GiB24.27 GiB127±37%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB1.13 GiB151.79 GiB26.77 GiB22±26.5%
dots.llm1.instMoEQ8_0143B141.37 GiB8.72 GiB151.02 GiB27.54 GiB75±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 32,768 context with q4_0 KV cache, the largest being MiMo-V2.5 at Q4_K_L. 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.