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. 2098 of 2118 indexed models fit at 128K context with f16 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 1802image 2audio tts 21audio asr 39embedding 26video 16

What fits at 128K context

largest quantization that fits, per model · 2098 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Step-3.7-FlashUD-Q5_K_S201B128.59 GiB49.03 GiB178.55 GiB0.01 GiB19±26.5%
MiniMax-M2.5MoEQ5_K_S229B146.66 GiB31.00 GiB178.55 GiB0.01 GiB42±37%
MiniMax-M2.1MoEI1-Q5_K_S229B146.66 GiB31.00 GiB178.55 GiB0.01 GiB42±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ5_K_S229B146.66 GiB31.00 GiB178.55 GiB0.01 GiB42±37%
MiMo-V2-FlashMoEKV unresolvedIQ4_NL310B162.41 GiB15.00 GiB178.36 GiB0.20 GiB61±37%
Trinity-Large-ThinkingMoEQ3_K_M399B168.90 GiB8.29 GiB178.12 GiB0.44 GiB83±37%
Trinity-Large-TrueBaseMoEQ3_K_M399B168.74 GiB8.29 GiB177.96 GiB0.60 GiB83±37%
Trinity-Large-PreviewMoEIQ3_M399B168.61 GiB8.29 GiB177.83 GiB0.73 GiB83±37%
Qwen3-Coder-REAP-363B-A35BMoEQ3_K_S363B145.81 GiB31.00 GiB177.74 GiB0.82 GiB38±37%
Qwen3.5-REAP-262B-A17BMoEQ5_K_M262B172.96 GiB3.75 GiB177.66 GiB0.90 GiB90±37%
dots.llm1.instMoEUD-IQ2_M143B52.59 GiB124.00 GiB177.52 GiB1.04 GiB15±37%
command-r-35b-writer-v2I1-Q3_K_M35.0B16.41 GiB160.00 GiB177.41 GiB1.15 GiB19±26.5%
grok-2MoEQ4_0270B144.10 GiB32.00 GiB177.14 GiB1.42 GiB25±37%
ERNIE-4.5-300B-A47B-PTIQ4_XS300B148.58 GiB27.00 GiB176.61 GiB1.95 GiB19±26.5%
step-3.5-flashQ5_K_S199B126.55 GiB49.03 GiB176.51 GiB2.05 GiB19±26.5%
Qwen2.5-72BF1672.7B135.44 GiB40.00 GiB176.47 GiB2.09 GiB19±26.5%
Kimi-Dev-72BBF1672.7B135.44 GiB40.00 GiB176.47 GiB2.09 GiB19±26.5%
Qwen2.5-VL-72B-InstructBF1673.4B135.44 GiB40.00 GiB176.47 GiB2.09 GiB19±26.5%
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16MoEUD-IQ1_M561B175.12 GiB0.00 GiB176.05 GiB2.51 GiB96±37%
GLM-5MoEUD-TQ1_0754B164.05 GiB10.97 GiB175.96 GiB2.60 GiB69±37%
GLM-5.1MoEIQ1_M754B163.93 GiB10.97 GiB175.85 GiB2.71 GiB69±37%
Noromaid-20b-v0.1.1Q8_020.0B19.79 GiB155.00 GiB175.73 GiB2.83 GiB19±26.5%
Nethena-20BQ8_020.0B19.79 GiB155.00 GiB175.73 GiB2.83 GiB19±26.5%
Hermes-4-405BIQ2_XS406B111.15 GiB63.00 GiB175.33 GiB3.23 GiB19±26.5%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB63.00 GiB175.33 GiB3.23 GiB19±26.5%
Qwen3-235B-A22B-Instruct-2507MoEQ5_K_S235B150.83 GiB23.50 GiB175.27 GiB3.29 GiB44±37%
Qwen3-235B-A22B-Thinking-2507MoEQ5_K_S235B150.83 GiB23.50 GiB175.27 GiB3.29 GiB44±37%
Qwen3-VL-235B-A22B-InstructMoEQ5_K_S236B150.76 GiB23.50 GiB175.20 GiB3.36 GiB44±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ5_K_S236B150.76 GiB23.50 GiB175.20 GiB3.36 GiB44±37%
Qwen3-235B-A22BMoEQ5_K_S235B150.76 GiB23.50 GiB175.20 GiB3.36 GiB44±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q5_K_S235B150.76 GiB23.50 GiB175.20 GiB3.36 GiB44±37%
GLM-4.7-REAP-218B-A32BMoEQ4_1218B128.16 GiB46.00 GiB175.10 GiB3.46 GiB30±37%
Qwen3.5-397B-A17BMoEUD-Q3_K_M403B170.04 GiB3.75 GiB174.74 GiB3.82 GiB98±37%
Hy3MoEIQ3_M299B133.80 GiB40.00 GiB174.74 GiB3.82 GiB35±37%
Ornith-1.0-397BMoEQ3_K_M397B169.03 GiB3.75 GiB173.73 GiB4.83 GiB99±37%
Nex-N2-ProMoEQ3_K_M397B169.03 GiB3.75 GiB173.73 GiB4.83 GiB99±37%
Apertus-70B-Instruct-2509BF1670.6B131.51 GiB40.00 GiB172.59 GiB5.97 GiB20±26.5%
Llama-3.3-70B-InstructF1670.6B131.43 GiB40.00 GiB172.45 GiB6.11 GiB20±26.5%
Hermes-4-70BBF1670.6B131.43 GiB40.00 GiB172.45 GiB6.11 GiB20±26.5%
Llama-3.1-70BF1670.6B131.43 GiB40.00 GiB172.45 GiB6.11 GiB20±26.5%
DeepSeek-R1-Distill-Llama-70BF1670.6B131.43 GiB40.00 GiB172.45 GiB6.11 GiB20±26.5%
Athene-70BBF1670.6B131.43 GiB40.00 GiB172.45 GiB6.11 GiB20±26.5%
Hermes-3-Llama-3.1-70BBF1670.6B131.43 GiB40.00 GiB172.45 GiB6.11 GiB20±26.5%
Meta-Llama-3-70B-Instruct-abliterated-v3.5BF1670.6B131.43 GiB40.00 GiB172.45 GiB6.11 GiB20±26.5%
L3.3-70B-Magnum-DiamondBF1670.6B131.43 GiB40.00 GiB172.45 GiB6.11 GiB20±26.5%
DeepSeek-V3.1-TerminusMoEIQ2_XXS685B162.59 GiB8.58 GiB172.15 GiB6.41 GiB76±37%
DeepSeek-V3.2MoEIQ2_XXS685B162.59 GiB8.58 GiB172.15 GiB6.41 GiB76±37%
cogito-671b-v2.1MoEIQ2_XXS671B162.59 GiB8.58 GiB172.15 GiB6.41 GiB76±37%
DeepSeek-V3-0324MoEIQ2_XXS685B162.45 GiB8.58 GiB172.01 GiB6.55 GiB76±37%
r1-1776MoEIQ2_XXS671B162.45 GiB8.58 GiB172.01 GiB6.55 GiB76±37%
DeepSeek-R1MoEIQ2_XXS685B162.45 GiB8.58 GiB172.01 GiB6.55 GiB76±37%
Qwen3-Coder-480B-A35B-InstructMoEUD-IQ1_M480B139.43 GiB31.00 GiB171.37 GiB7.19 GiB40±37%
MiMo-V2.5MoEKV unresolvedIQ4_XS311B154.32 GiB15.00 GiB170.27 GiB8.29 GiB62±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ8_0139B137.78 GiB31.00 GiB169.67 GiB8.89 GiB40±37%
GLM-4.5MoEQ2_K_L358B122.32 GiB46.00 GiB169.26 GiB9.30 GiB32±37%
GLM-4.7MoEQ2_K_L358B122.32 GiB46.00 GiB169.26 GiB9.30 GiB32±37%
Devstral-2-123B-Instruct-2512Q8_0125B123.73 GiB44.00 GiB168.79 GiB9.77 GiB20±26.5%
GLM-4.6MoEQ2_K_L357B121.85 GiB46.00 GiB168.79 GiB9.77 GiB32±37%
Mistral-Medium-3.5-128BQ8_0128B123.73 GiB44.00 GiB168.79 GiB9.77 GiB20±26.5%
DeepSeek-V3.1MoEUD-TQ1_0685B158.79 GiB8.58 GiB168.35 GiB10.21 GiB77±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?
2098 of 2118 indexed open-weight models fit a Instinct MI300X at 131,072 context with f16 KV cache, the largest being Step-3.7-Flash at UD-Q5_K_S. 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.