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 16K 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 16K context

largest quantization that fits, per model · 2108 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-Coder-480B-A35B-InstructMoEIQ3_XXS480B176.37 GiB1.09 GiB178.39 GiB0.17 GiB86±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.13 GiB178.02 GiB0.54 GiB116±37%
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.26 GiB177.96 GiB0.60 GiB81±37%
DeepSeek-R1-0528MoEIQ2_S685B176.61 GiB0.30 GiB177.89 GiB0.67 GiB103±37%
DeepSeek-V3.1-TerminusMoEIQ2_S685B176.61 GiB0.30 GiB177.89 GiB0.67 GiB103±37%
DeepSeek-V3.1MoEIQ2_S685B176.61 GiB0.30 GiB177.89 GiB0.67 GiB103±37%
Trinity-Large-PreviewMoEQ3_K_M399B176.43 GiB0.49 GiB177.84 GiB0.72 GiB119±37%
Trinity-Large-TrueBaseMoEI1-Q3_K_M399B176.30 GiB0.49 GiB177.72 GiB0.84 GiB119±37%
MiMo-V2.5MoEKV unresolvedQ4_K_L311B176.24 GiB0.53 GiB177.71 GiB0.85 GiB102±37%
Trinity-Large-ThinkingMoEIQ3_M399B176.24 GiB0.49 GiB177.65 GiB0.91 GiB119±37%
Ornith-1.0-397BMoEIQ3_M397B176.54 GiB0.13 GiB177.62 GiB0.94 GiB116±37%
Hy3MoEQ4_1299B174.90 GiB1.41 GiB177.24 GiB1.32 GiB91±37%
MiniMax-M2.7MoEUD-Q6_K229B175.15 GiB1.09 GiB177.12 GiB1.44 GiB100±37%
MiniMax-M2.1MoEQ6_K229B174.91 GiB1.09 GiB176.89 GiB1.67 GiB100±37%
MiniMax-M2MoEQ6_K229B174.91 GiB1.09 GiB176.89 GiB1.67 GiB100±37%
MiniMax-M2.5MoEQ6_K229B174.87 GiB1.09 GiB176.84 GiB1.72 GiB100±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ6_K229B174.87 GiB1.09 GiB176.84 GiB1.72 GiB100±37%
ERNIE-4.5-300B-A47B-PTQ4_1300B174.57 GiB0.95 GiB176.54 GiB2.02 GiB19±26.5%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB0.13 GiB176.37 GiB2.19 GiB117±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_L310B174.72 GiB0.53 GiB176.19 GiB2.37 GiB103±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.05 GiB2.51 GiB96±37%
MiniMax-M3MoEQ3_K_S427B174.39 GiB0.53 GiB175.83 GiB2.73 GiB103±37%
grok-2MoEQ5_K_S270B173.10 GiB1.13 GiB175.27 GiB3.29 GiB36±37%
GLM-4.6-REAP-268B-A32BMoEQ5_K_S269B172.64 GiB1.62 GiB175.20 GiB3.36 GiB76±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB0.30 GiB174.73 GiB3.83 GiB105±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB0.30 GiB174.40 GiB4.16 GiB105±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB0.30 GiB174.39 GiB4.17 GiB105±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.13 GiB174.04 GiB4.52 GiB107±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB0.30 GiB173.67 GiB4.89 GiB106±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB0.30 GiB173.38 GiB5.18 GiB106±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB0.30 GiB172.76 GiB5.80 GiB106±37%
Hermes-3-Llama-3.1-405BIQ3_M406B169.26 GiB2.21 GiB172.66 GiB5.90 GiB20±26.5%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB0.53 GiB172.50 GiB6.06 GiB105±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB0.39 GiB170.66 GiB7.90 GiB107±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB1.62 GiB170.12 GiB8.44 GiB71±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB0.84 GiB167.42 GiB11.14 GiB112±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB0.39 GiB165.38 GiB13.18 GiB109±37%
GLM-5.1MoEIQ1_M754B163.93 GiB0.39 GiB165.27 GiB13.29 GiB109±37%
r1-1776MoEIQ2_XXS671B162.45 GiB0.30 GiB163.73 GiB14.83 GiB111±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB0.30 GiB163.73 GiB14.83 GiB111±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB1.09 GiB163.60 GiB14.96 GiB83±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB1.98 GiB163.11 GiB15.45 GiB21±26.5%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB0.84 GiB162.57 GiB15.99 GiB135±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB1.62 GiB162.06 GiB16.50 GiB91±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB1.62 GiB162.06 GiB16.50 GiB91±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB1.62 GiB161.44 GiB17.12 GiB91±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB1.62 GiB160.81 GiB17.75 GiB92±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB0.83 GiB157.19 GiB21.37 GiB89±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB0.83 GiB157.19 GiB21.37 GiB89±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB0.83 GiB157.12 GiB21.44 GiB89±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB0.83 GiB157.12 GiB21.44 GiB89±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB0.83 GiB157.12 GiB21.44 GiB89±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB0.83 GiB157.12 GiB21.44 GiB89±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB0.30 GiB156.97 GiB21.59 GiB109±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB0.30 GiB156.97 GiB21.59 GiB109±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB0.30 GiB156.97 GiB21.59 GiB109±37%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.02 GiB154.29 GiB24.27 GiB127±37%
step-3.5-flashQ6_K199B150.81 GiB1.98 GiB153.72 GiB24.84 GiB22±26.5%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB0.56 GiB151.22 GiB27.34 GiB22±26.5%
Qwen3-Coder-NextMoEBF1679.7B148.51 GiB0.42 GiB149.82 GiB28.74 GiB131±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 16,384 context with q4_0 KV cache, the largest being Qwen3-Coder-480B-A35B-Instruct at IQ3_XXS. 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.