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 8K 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
vision language 192text 1812image 2audio tts 21audio asr 39video 16embedding 26

What fits at 8K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.07 GiB177.95 GiB0.61 GiB117±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.19 GiB177.89 GiB0.67 GiB82±37%
Qwen3-Coder-480B-A35B-InstructMoEIQ3_XXS480B176.37 GiB0.54 GiB177.85 GiB0.71 GiB88±37%
DeepSeek-R1-0528MoEIQ2_S685B176.61 GiB0.15 GiB177.74 GiB0.82 GiB104±37%
DeepSeek-V3.1-TerminusMoEIQ2_S685B176.61 GiB0.15 GiB177.74 GiB0.82 GiB104±37%
DeepSeek-V3.1MoEIQ2_S685B176.61 GiB0.15 GiB177.74 GiB0.82 GiB104±37%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB0.35 GiB177.71 GiB0.85 GiB119±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB0.35 GiB177.59 GiB0.97 GiB120±37%
Ornith-1.0-397BMoEIQ3_M397B176.54 GiB0.07 GiB177.56 GiB1.00 GiB117±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB0.35 GiB177.52 GiB1.04 GiB120±37%
MiMo-V2.5MoEKV unresolvedQ4_K_L311B176.24 GiB0.26 GiB177.45 GiB1.11 GiB104±37%
MiniMax-M2.7MoEUD-Q6_K229B175.15 GiB0.54 GiB176.58 GiB1.98 GiB103±37%
Hy3MoEQ4_1299B174.90 GiB0.70 GiB176.54 GiB2.02 GiB94±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB0.54 GiB176.34 GiB2.22 GiB103±37%
MiniMax-M2MoEQ6_K229B174.91 GiB0.54 GiB176.34 GiB2.22 GiB103±37%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB0.07 GiB176.30 GiB2.26 GiB118±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB0.54 GiB176.30 GiB2.26 GiB103±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB0.54 GiB176.30 GiB2.26 GiB103±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB0.47 GiB176.07 GiB2.49 GiB19±26.5%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.05 GiB2.51 GiB96±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_L310B174.72 GiB0.26 GiB175.93 GiB2.63 GiB105±37%
MiniMax-M3MoEQ3_K_S427B174.39 GiB0.26 GiB175.57 GiB2.99 GiB105±37%
grok-2MoEQ5_K_S270B173.10 GiB0.56 GiB174.71 GiB3.85 GiB36±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB0.15 GiB174.58 GiB3.98 GiB106±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_S269B172.64 GiB0.81 GiB174.39 GiB4.17 GiB79±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB0.15 GiB174.25 GiB4.31 GiB106±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB0.15 GiB174.24 GiB4.32 GiB106±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.07 GiB173.98 GiB4.58 GiB108±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB0.15 GiB173.52 GiB5.04 GiB106±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB0.15 GiB173.23 GiB5.33 GiB107±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB0.15 GiB172.61 GiB5.95 GiB107±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB0.26 GiB172.24 GiB6.32 GiB106±37%
Hermes-3-Llama-3.1-405BIQ3_M406B169.26 GiB1.11 GiB171.55 GiB7.01 GiB20±26.5%
GLM-5.2MoEQ3_K_M753B169.33 GiB0.19 GiB170.47 GiB8.09 GiB108±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB0.81 GiB169.32 GiB9.24 GiB73±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB0.42 GiB167.00 GiB11.56 GiB114±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB0.19 GiB165.19 GiB13.37 GiB111±37%
GLM-5.1MoEIQ1_M754B163.93 GiB0.19 GiB165.07 GiB13.49 GiB111±37%
r1-1776MoEIQ2_XXS671B162.45 GiB0.15 GiB163.58 GiB14.98 GiB112±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB0.15 GiB163.58 GiB14.98 GiB112±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB0.54 GiB163.05 GiB15.51 GiB84±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB1.13 GiB162.26 GiB16.30 GiB21±26.5%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB0.42 GiB162.15 GiB16.41 GiB139±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB0.81 GiB161.26 GiB17.30 GiB95±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB0.81 GiB161.26 GiB17.30 GiB95±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB0.81 GiB160.63 GiB17.93 GiB95±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB0.81 GiB160.00 GiB18.56 GiB95±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB0.15 GiB156.83 GiB21.73 GiB110±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB0.15 GiB156.83 GiB21.73 GiB110±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB0.15 GiB156.83 GiB21.73 GiB110±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB0.41 GiB156.77 GiB21.79 GiB90±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB0.41 GiB156.77 GiB21.79 GiB90±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB0.41 GiB156.70 GiB21.86 GiB90±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB0.41 GiB156.70 GiB21.86 GiB90±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB0.41 GiB156.70 GiB21.86 GiB90±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB0.41 GiB156.70 GiB21.86 GiB90±37%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.02 GiB154.29 GiB24.27 GiB127±37%
step-3.5-flashQ6_K199B150.81 GiB1.13 GiB152.88 GiB25.68 GiB22±26.5%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB0.28 GiB150.94 GiB27.62 GiB22±26.5%
Qwen3-Coder-NextMoEBF1679.7B148.51 GiB0.21 GiB149.61 GiB28.95 GiB133±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 8,192 context with q4_0 KV cache, the largest being Nex-N2-Pro 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.