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 4K 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 4K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.7-REAP-218B-A32BMoEQ4_0218B117.58 GiB0.40 GiB118.92 GiB0.12 GiB66±37%
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.03 GiB118.79 GiB0.25 GiB110±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_XS310B117.68 GiB0.13 GiB118.76 GiB0.28 GiB98±37%
Hermes-4-405BIQ2_S406B117.02 GiB0.55 GiB118.75 GiB0.29 GiB18±26.5%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.02 GiB118.74 GiB0.30 GiB98±37%
Hy3MoEIQ3_XXS299B117.43 GiB0.35 GiB118.72 GiB0.32 GiB88±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.14 GiB118.46 GiB0.58 GiB77±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ4_XS235B117.24 GiB0.21 GiB118.38 GiB0.66 GiB76±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_XS235B117.24 GiB0.21 GiB118.38 GiB0.66 GiB76±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_XXS311B117.27 GiB0.13 GiB118.35 GiB0.69 GiB98±37%
Step-3.7-FlashQ4_1201B116.67 GiB0.71 GiB118.31 GiB0.73 GiB18±26.5%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB0.21 GiB118.02 GiB1.02 GiB77±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.07 GiB117.95 GiB1.09 GiB93±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.07 GiB117.95 GiB1.09 GiB93±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.07 GiB117.95 GiB1.09 GiB93±37%
Qwen3-VL-235B-A22B-ThinkingMoEIQ4_XS236B116.70 GiB0.21 GiB117.84 GiB1.20 GiB77±37%
Qwen3-VL-235B-A22B-InstructMoEIQ4_XS236B116.70 GiB0.21 GiB117.84 GiB1.20 GiB77±37%
ERNIE-4.5-300B-A47B-PTUD-IQ3_XXS300B116.56 GiB0.24 GiB117.82 GiB1.22 GiB18±26.5%
Qwen3-235B-A22B-abliteratedMoEI1-IQ4_XS235B116.68 GiB0.21 GiB117.82 GiB1.22 GiB77±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.26 GiB117.69 GiB1.35 GiB113±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.26 GiB117.69 GiB1.35 GiB113±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.03 GiB117.40 GiB1.64 GiB100±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.03 GiB116.81 GiB2.23 GiB111±37%
step-3.5-flashQ4_1199B115.15 GiB0.71 GiB116.79 GiB2.25 GiB18±26.5%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.05 GiB116.70 GiB2.34 GiB100±37%
grok-2MoEIQ3_M270B115.25 GiB0.28 GiB116.57 GiB2.47 GiB34±37%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB0.40 GiB115.38 GiB3.66 GiB84±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB0.40 GiB115.38 GiB3.66 GiB84±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB0.27 GiB115.15 GiB3.89 GiB99±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB0.27 GiB114.93 GiB4.11 GiB99±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB0.27 GiB114.93 GiB4.11 GiB99±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB0.40 GiB114.90 GiB4.14 GiB85±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB0.27 GiB114.69 GiB4.35 GiB99±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB0.21 GiB113.61 GiB5.43 GiB127±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.26 GiB113.22 GiB5.82 GiB116±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.03 GiB113.22 GiB5.82 GiB114±37%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB0.55 GiB112.89 GiB6.15 GiB19±26.5%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB0.21 GiB112.77 GiB6.27 GiB107±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB0.27 GiB111.38 GiB7.66 GiB102±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB0.20 GiB110.52 GiB8.52 GiB81±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB0.20 GiB110.52 GiB8.52 GiB81±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.02 GiB110.22 GiB8.82 GiB113±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB0.40 GiB109.82 GiB9.22 GiB78±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB0.13 GiB109.65 GiB9.39 GiB104±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.02 GiB109.06 GiB9.98 GiB114±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB0.27 GiB109.06 GiB9.98 GiB79±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB0.25 GiB108.80 GiB10.24 GiB36±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB0.40 GiB108.49 GiB10.55 GiB89±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.10 GiB107.86 GiB11.18 GiB96±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB0.21 GiB107.80 GiB11.24 GiB83±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB0.27 GiB107.56 GiB11.48 GiB89±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB0.27 GiB107.56 GiB11.48 GiB89±37%
GLM-4.6VMoEQ8_0108B105.81 GiB0.20 GiB106.94 GiB12.10 GiB83±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.03 GiB106.38 GiB12.66 GiB102±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB0.51 GiB104.47 GiB14.57 GiB20±26.5%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB0.28 GiB104.10 GiB14.94 GiB20±26.5%
dots.llm1.instMoEQ5_K_M143B101.84 GiB1.09 GiB103.86 GiB15.18 GiB90±37%
gpt-oss-20b-hereticMoEQ8_020.9B102.72 GiB0.03 GiB103.64 GiB15.40 GiB61±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ8_020.9B102.72 GiB0.03 GiB103.64 GiB15.40 GiB61±37%
gemma-2-27b-itF3227.2B101.43 GiB0.40 GiB102.87 GiB16.17 GiB20±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
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 4,096 context with q4_0 KV cache, the largest being GLM-4.7-REAP-218B-A32B at Q4_0. 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.