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 32K context with q8_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 32K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB1.12 GiB119.00 GiB0.04 GiB86±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB1.12 GiB119.00 GiB0.04 GiB86±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB1.12 GiB119.00 GiB0.04 GiB86±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB4.12 GiB119.00 GiB0.04 GiB75±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB1.42 GiB118.85 GiB0.19 GiB102±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB1.42 GiB118.85 GiB0.19 GiB102±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB4.12 GiB118.78 GiB0.26 GiB76±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB4.12 GiB118.78 GiB0.26 GiB76±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB4.12 GiB118.53 GiB0.51 GiB76±37%
command-a-plus-05-2026-bf16MoEIQ4_NL219B116.79 GiB0.76 GiB118.45 GiB0.59 GiB74±37%
step-3.5-flashQ4_K199B110.56 GiB6.92 GiB118.41 GiB0.63 GiB18±26.5%
Qwen3.5-397B-A17BMoEIQ2_XS403B116.79 GiB0.50 GiB118.24 GiB0.80 GiB106±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.50 GiB117.87 GiB1.17 GiB97±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_L235B113.46 GiB3.12 GiB117.51 GiB1.53 GiB66±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.50 GiB117.28 GiB1.76 GiB107±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.61 GiB117.27 GiB1.77 GiB95±37%
GLM-4.7-REAP-218B-A32BMoEIQ4_XS218B110.09 GiB6.11 GiB117.14 GiB1.90 GiB53±37%
Step-3.7-FlashQ4_K_S201B109.05 GiB6.92 GiB116.90 GiB2.14 GiB18±26.5%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB3.19 GiB116.59 GiB2.45 GiB97±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_XXS310B113.14 GiB1.99 GiB116.08 GiB2.96 GiB87±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB3.19 GiB115.75 GiB3.29 GiB85±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB6.11 GiB115.52 GiB3.52 GiB57±37%
grok-2MoEQ3_K_S270B109.94 GiB4.25 GiB115.23 GiB3.81 GiB32±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB4.12 GiB115.23 GiB3.81 GiB77±37%
GLM-4.5MoEUD-IQ2_XXS358B107.95 GiB6.11 GiB115.00 GiB4.04 GiB62±37%
GLM-4.7MoEUD-IQ2_XXS358B107.95 GiB6.11 GiB115.00 GiB4.04 GiB62±37%
GLM-4.6MoEUD-IQ2_XXS357B107.48 GiB6.11 GiB114.53 GiB4.51 GiB63±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB1.42 GiB114.37 GiB4.67 GiB105±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB6.11 GiB114.19 GiB4.85 GiB63±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.50 GiB113.68 GiB5.36 GiB110±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB3.05 GiB113.37 GiB5.67 GiB68±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB3.05 GiB113.37 GiB5.67 GiB68±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB4.12 GiB112.90 GiB6.14 GiB64±37%
dots.llm1.instMoEQ5_K_S143B95.08 GiB16.47 GiB112.48 GiB6.56 GiB43±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB3.72 GiB112.27 GiB6.77 GiB33±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB1.99 GiB111.51 GiB7.53 GiB90±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB4.12 GiB111.41 GiB7.63 GiB69±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB4.12 GiB111.41 GiB7.63 GiB69±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB3.19 GiB110.78 GiB8.26 GiB69±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.03 GiB110.24 GiB8.80 GiB113±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB1.99 GiB109.92 GiB9.12 GiB91±37%
GLM-4.6VMoEQ8_0108B105.81 GiB3.05 GiB109.79 GiB9.25 GiB70±37%
Hermes-4-405BIQ2_XXS406B99.91 GiB8.37 GiB109.45 GiB9.59 GiB19±26.5%
Hermes-3-Llama-3.1-405BIQ2_XXS406B99.91 GiB8.37 GiB109.45 GiB9.59 GiB19±26.5%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB1.46 GiB109.23 GiB9.81 GiB87±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.03 GiB109.08 GiB9.96 GiB114±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB3.12 GiB108.78 GiB10.26 GiB70±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB3.12 GiB108.78 GiB10.26 GiB70±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB3.12 GiB108.78 GiB10.26 GiB70±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB3.12 GiB108.78 GiB10.26 GiB70±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB3.12 GiB108.78 GiB10.26 GiB70±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB4.25 GiB108.07 GiB10.97 GiB19±26.5%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB5.84 GiB108.03 GiB11.01 GiB19±26.5%
Hy3MoEQ2_K299B101.28 GiB5.31 GiB107.53 GiB11.51 GiB71±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB3.28 GiB107.24 GiB11.80 GiB20±26.5%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.50 GiB106.84 GiB12.20 GiB99±37%
ERNIE-4.5-300B-A47B-PTQ2_K_L300B101.76 GiB3.59 GiB106.37 GiB12.67 GiB20±26.5%
gemma-2-27b-itF3227.2B101.43 GiB3.48 GiB105.95 GiB13.09 GiB20±26.5%
gpt-oss-20b-hereticMoEQ8_020.9B102.72 GiB0.41 GiB104.02 GiB15.02 GiB59±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ8_020.9B102.72 GiB0.41 GiB104.02 GiB15.02 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 32,768 context with q8_0 KV cache, the largest being DeepSeek-Coder-V2-Instruct-0724 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.