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 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 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◐
command-a-plus-05-2026-bf16MoEQ6_K219B176.79 GiB0.76 GiB178.45 GiB0.11 GiB80±37%
Trinity-Large-ThinkingMoEQ3_K_L399B176.09 GiB1.42 GiB178.43 GiB0.13 GiB112±37%
grok-2MoEQ5_K_S270B173.10 GiB4.25 GiB178.39 GiB0.17 GiB34±37%
Nex-N2-ProMoEIQ3_M397B176.93 GiB0.50 GiB178.38 GiB0.18 GiB114±37%
Trinity-Large-PreviewMoEQ3_K_L399B175.82 GiB1.42 GiB178.16 GiB0.40 GiB113±37%
Trinity-Large-TrueBaseMoEQ3_K_L399B175.82 GiB1.42 GiB178.16 GiB0.40 GiB113±37%
Ornith-1.0-397BMoEIQ3_M397B176.54 GiB0.50 GiB177.99 GiB0.57 GiB114±37%
MiMo-V2-FlashMoEKV unresolvedQ4_K_L310B174.72 GiB1.99 GiB177.66 GiB0.90 GiB96±37%
DeepSeek-R1-0528MoEIQ2_XS685B175.47 GiB1.14 GiB177.59 GiB0.97 GiB100±37%
DeepSeek-V3.1MoEIQ2_XS685B175.47 GiB1.14 GiB177.59 GiB0.97 GiB100±37%
MiniMax-M3MoEQ3_K_S427B174.39 GiB1.99 GiB177.30 GiB1.26 GiB96±37%
Qwen3.5-397B-A17BMoEQ3_K_M403B175.29 GiB0.50 GiB176.73 GiB1.83 GiB115±37%
Hy3MoEQ4_K_L299B169.99 GiB5.31 GiB176.24 GiB2.32 GiB78±37%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.05 GiB2.51 GiB96±37%
DeepSeek-V3.1-TerminusMoEUD-IQ1_S685B173.84 GiB1.14 GiB175.96 GiB2.60 GiB100±37%
DeepSeek-V3-0324MoEUD-IQ1_S685B173.45 GiB1.14 GiB175.57 GiB2.99 GiB101±37%
cogito-v2-preview-deepseek-671B-MoEMoEUD-IQ1_S671B173.12 GiB1.14 GiB175.24 GiB3.32 GiB101±37%
cogito-671b-v2.1MoEUD-IQ1_S671B173.11 GiB1.14 GiB175.23 GiB3.33 GiB101±37%
GLM-4.7-REAP-218B-A32BMoEQ6_K218B167.57 GiB6.11 GiB174.62 GiB3.94 GiB61±37%
DeepSeek-TNG-R1T2-ChimeraMoEUD-IQ1_S685B172.39 GiB1.14 GiB174.51 GiB4.05 GiB101±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB0.50 GiB174.41 GiB4.15 GiB105±37%
DeepSeek-Prover-V2-671BMoEUD-IQ1_S685B172.10 GiB1.14 GiB174.22 GiB4.34 GiB101±37%
Minimax-M3-abliterated-cleanMoEQ3_K_S427B171.06 GiB1.99 GiB173.97 GiB4.59 GiB97±37%
DeepSeek-V3.2MoEUD-IQ1_S685B171.48 GiB1.14 GiB173.60 GiB4.96 GiB101±37%
Hermes-3-Llama-3.1-405BQ3_K_S406B163.20 GiB8.37 GiB172.75 GiB5.81 GiB20±26.5%
ERNIE-4.5-300B-A47B-PTQ4_K_M300B167.78 GiB3.59 GiB172.39 GiB6.17 GiB20±26.5%
MiMo-V2.5MoEKV unresolvedQ4_K_S311B169.41 GiB1.99 GiB172.35 GiB6.21 GiB98±37%
GLM-5.2MoEQ3_K_M753B169.33 GiB1.46 GiB171.73 GiB6.83 GiB101±37%
Solar-Open2-250BMoEQ5_K_M250B165.65 GiB3.19 GiB169.76 GiB8.80 GiB99±37%
Step-3.7-FlashQ6_K_L201B160.20 GiB6.92 GiB168.05 GiB10.51 GiB20±26.5%
Qwen3-Coder-480B-A35B-InstructMoEQ2_K_L480B162.87 GiB4.12 GiB167.92 GiB10.64 GiB80±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_M363B161.57 GiB4.12 GiB166.62 GiB11.94 GiB73±37%
GLM-4.5MoEQ3_K_M358B159.51 GiB6.11 GiB166.56 GiB12.00 GiB75±37%
GLM-4.7MoEQ3_K_M358B159.51 GiB6.11 GiB166.56 GiB12.00 GiB75±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB1.46 GiB166.45 GiB12.11 GiB103±37%
GLM-5.1MoEIQ1_M754B163.93 GiB1.46 GiB166.34 GiB12.22 GiB103±37%
GLM-4.6MoEQ3_K_M357B158.89 GiB6.11 GiB165.94 GiB12.62 GiB76±37%
GLM-4.6-Derestricted-v3MoEQ3_K_L357B158.25 GiB6.11 GiB165.30 GiB13.26 GiB76±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedQ3_K_S402B160.80 GiB3.19 GiB164.91 GiB13.65 GiB116±37%
r1-1776MoEIQ2_XXS671B162.45 GiB1.14 GiB164.57 GiB13.99 GiB106±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB1.14 GiB164.57 GiB13.99 GiB106±37%
GLM-4.6-REAP-268B-A32BMoEQ4_1269B156.99 GiB6.11 GiB164.04 GiB14.52 GiB70±37%
MiniMax-M2.7MoEUD-Q5_K_M229B157.23 GiB4.12 GiB162.24 GiB16.32 GiB93±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_M235B155.43 GiB3.12 GiB159.48 GiB19.08 GiB80±37%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_M235B155.43 GiB3.12 GiB159.48 GiB19.08 GiB80±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_M236B155.36 GiB3.12 GiB159.41 GiB19.15 GiB80±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_M236B155.36 GiB3.12 GiB159.41 GiB19.15 GiB80±37%
Qwen3-235B-A22BMoEQ5_K_M235B155.36 GiB3.12 GiB159.41 GiB19.15 GiB80±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_M235B155.36 GiB3.12 GiB159.41 GiB19.15 GiB80±37%
dots.llm1.instMoEQ8_0143B141.37 GiB16.47 GiB158.77 GiB19.79 GiB58±37%
step-3.5-flashQ6_K199B150.81 GiB6.92 GiB158.66 GiB19.90 GiB21±26.5%
DeepSeek-Coder-V2-Instruct-0724MoEQ5_K236B155.74 GiB1.12 GiB157.80 GiB20.76 GiB104±37%
DeepSeek-V2.5MoEQ5_K236B155.74 GiB1.12 GiB157.80 GiB20.76 GiB104±37%
DeepSeek-Coder-V2-InstructMoEQ5_K_M236B155.74 GiB1.12 GiB157.80 GiB20.76 GiB104±37%
MiniMax-M2.1MoEQ5_K_M229B151.23 GiB4.12 GiB156.23 GiB22.33 GiB96±37%
MiniMax-M2MoEQ5_K_M229B151.23 GiB4.12 GiB156.23 GiB22.33 GiB96±37%
MiniMax-M2.5MoEQ5_K_M229B151.16 GiB4.12 GiB156.16 GiB22.40 GiB96±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ5_K_M229B151.16 GiB4.12 GiB156.16 GiB22.40 GiB96±37%
DeepSeek-V4-FlashMoEQ4_K291B153.33 GiB0.03 GiB154.31 GiB24.25 GiB127±37%
Hunyuan-A13B-InstructMoEBF1680.4B149.76 GiB2.13 GiB152.79 GiB25.77 GiB22±26.5%
●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 q8_0 KV cache, the largest being command-a-plus-05-2026-bf16 at Q6_K. 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.