AMD · datacenter

Instinct MI250X

Instinct MI250X has 128 GB of VRAM at 3277 GB/s — about 119.04 GiB usable after driver and compositor overhead. 2091 of 2118 indexed models fit at 64K context with q4_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
128 GB
HBM2e
Bandwidth
3277 GB/s
8192-bit bus
Tensor FP16
383 TF
dense
TDP
560 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1796vision language 191audio tts 21image 2audio asr 39video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB4.36 GiB119.02 GiB0.02 GiB74±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB4.36 GiB119.02 GiB0.02 GiB74±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.68 GiB119.00 GiB0.04 GiB74±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB4.36 GiB118.77 GiB0.27 GiB75±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB1.28 GiB118.71 GiB0.33 GiB103±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB1.28 GiB118.71 GiB0.33 GiB103±37%
step-3.5-flashQ4_K199B110.56 GiB7.04 GiB118.53 GiB0.51 GiB18±26.5%
Qwen3.5-397B-A17BMoEIQ2_XS403B116.79 GiB0.53 GiB118.27 GiB0.77 GiB106±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.53 GiB117.90 GiB1.14 GiB96±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_L235B113.46 GiB3.30 GiB117.70 GiB1.34 GiB65±37%
GLM-4.7-REAP-218B-A32BMoEIQ4_XS218B110.09 GiB6.47 GiB117.50 GiB1.54 GiB52±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.53 GiB117.31 GiB1.73 GiB107±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.64 GiB117.29 GiB1.75 GiB95±37%
Step-3.7-FlashQ4_K_S201B109.05 GiB7.04 GiB117.02 GiB2.02 GiB18±26.5%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB3.38 GiB116.78 GiB2.26 GiB95±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_XXS310B113.14 GiB2.11 GiB116.20 GiB2.84 GiB87±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ3_K_L236B113.97 GiB1.19 GiB116.09 GiB2.95 GiB88±37%
DeepSeek-V2.5MoEQ3_K_L236B113.97 GiB1.19 GiB116.09 GiB2.95 GiB88±37%
DeepSeek-Coder-V2-InstructMoEQ3_K_L236B113.97 GiB1.19 GiB116.09 GiB2.95 GiB88±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB3.38 GiB115.93 GiB3.11 GiB84±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB6.47 GiB115.88 GiB3.16 GiB56±37%
grok-2MoEQ3_K_S270B109.94 GiB4.50 GiB115.48 GiB3.56 GiB31±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB4.36 GiB115.47 GiB3.57 GiB76±37%
GLM-4.5MoEUD-IQ2_XXS358B107.95 GiB6.47 GiB115.36 GiB3.68 GiB61±37%
GLM-4.7MoEUD-IQ2_XXS358B107.95 GiB6.47 GiB115.36 GiB3.68 GiB61±37%
GLM-4.6MoEUD-IQ2_XXS357B107.48 GiB6.47 GiB114.89 GiB4.15 GiB62±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB6.47 GiB114.55 GiB4.49 GiB62±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB1.28 GiB114.23 GiB4.81 GiB106±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.53 GiB113.71 GiB5.33 GiB109±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB3.23 GiB113.55 GiB5.49 GiB67±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB3.23 GiB113.55 GiB5.49 GiB67±37%
dots.llm1.instMoEQ5_K_S143B95.08 GiB17.44 GiB113.45 GiB5.59 GiB42±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB4.36 GiB113.14 GiB5.90 GiB63±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB3.94 GiB112.49 GiB6.55 GiB32±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB4.36 GiB111.65 GiB7.39 GiB69±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB4.36 GiB111.65 GiB7.39 GiB69±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB2.11 GiB111.63 GiB7.41 GiB89±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB3.38 GiB110.97 GiB8.07 GiB68±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.02 GiB110.22 GiB8.82 GiB113±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB2.11 GiB110.03 GiB9.01 GiB90±37%
GLM-4.6VMoEQ8_0108B105.81 GiB3.23 GiB109.97 GiB9.07 GiB69±37%
Hermes-4-405BIQ2_XXS406B99.91 GiB8.86 GiB109.95 GiB9.09 GiB19±26.5%
Hermes-3-Llama-3.1-405BIQ2_XXS406B99.91 GiB8.86 GiB109.95 GiB9.09 GiB19±26.5%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB1.55 GiB109.31 GiB9.73 GiB86±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.02 GiB109.06 GiB9.98 GiB114±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB3.30 GiB108.96 GiB10.08 GiB69±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB3.30 GiB108.96 GiB10.08 GiB69±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB3.30 GiB108.96 GiB10.08 GiB69±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB3.30 GiB108.96 GiB10.08 GiB69±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB3.30 GiB108.96 GiB10.08 GiB69±37%
MiniMax-M2.7MoEUD-IQ4_NL229B103.15 GiB4.36 GiB108.39 GiB10.65 GiB79±37%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB6.19 GiB108.37 GiB10.67 GiB19±26.5%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB4.50 GiB108.32 GiB10.72 GiB19±26.5%
Hy3MoEQ2_K299B101.28 GiB5.63 GiB107.84 GiB11.20 GiB70±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB3.14 GiB107.10 GiB11.94 GiB20±26.5%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.53 GiB106.87 GiB12.17 GiB98±37%
ERNIE-4.5-300B-A47B-PTQ2_K_L300B101.76 GiB3.80 GiB106.58 GiB12.46 GiB20±26.5%
gemma-2-27b-itF3227.2B101.43 GiB3.46 GiB105.93 GiB13.11 GiB20±26.5%
gpt-oss-20b-hereticMoEQ8_020.9B102.72 GiB0.43 GiB104.04 GiB15.00 GiB59±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ8_020.9B102.72 GiB0.43 GiB104.04 GiB15.00 GiB59±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
Image generation11.55 it/s8.33–15.626
Benchmarked· n=6

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 vladmandic-sd-data-benchmark, which publishes no licence — so we display and link rather than redistribute them.

Questions people ask

What AI models can a Instinct MI250X run?
2091 of 2118 indexed open-weight models fit a Instinct MI250X at 65,536 context with q4_0 KV cache, the largest being MiniMax-M2.1 at IQ4_XS. That covers text, vision-language, image, video and speech models.
How much usable memory does a Instinct MI250X actually have?
Its nameplate is 128 GB, but about 119.04 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Instinct MI250X fast for local AI?
Its memory bandwidth is 3277 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.