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 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 4K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Hy3MoEIQ3_XXS299B117.43 GiB0.66 GiB119.03 GiB0.01 GiB86±37%
Step-3.7-FlashQ4_1201B116.67 GiB1.34 GiB118.95 GiB0.09 GiB18±26.5%
MiMo-V2-FlashMoEKV unresolvedIQ3_XS310B117.68 GiB0.25 GiB118.88 GiB0.16 GiB97±37%
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.06 GiB118.82 GiB0.22 GiB110±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.05 GiB118.77 GiB0.27 GiB98±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.27 GiB118.59 GiB0.45 GiB76±37%
Qwen3-235B-A22B-Instruct-2507MoEIQ4_XS235B117.24 GiB0.39 GiB118.56 GiB0.48 GiB75±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_XS235B117.24 GiB0.39 GiB118.56 GiB0.48 GiB75±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_XXS311B117.27 GiB0.25 GiB118.47 GiB0.57 GiB97±37%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB0.39 GiB118.21 GiB0.83 GiB76±37%
ERNIE-4.5-300B-A47B-PTUD-IQ3_XXS300B116.56 GiB0.45 GiB118.03 GiB1.01 GiB18±26.5%
Qwen3-VL-235B-A22B-ThinkingMoEIQ4_XS236B116.70 GiB0.39 GiB118.02 GiB1.02 GiB76±37%
Qwen3-VL-235B-A22B-InstructMoEIQ4_XS236B116.70 GiB0.39 GiB118.02 GiB1.02 GiB76±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.14 GiB118.02 GiB1.02 GiB93±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.14 GiB118.02 GiB1.02 GiB93±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.14 GiB118.02 GiB1.02 GiB93±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ4_XS235B116.68 GiB0.39 GiB118.00 GiB1.04 GiB76±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.50 GiB117.93 GiB1.11 GiB110±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.50 GiB117.93 GiB1.11 GiB110±37%
GLM-4.7-REAP-218B-A32BMoEIQ4_NL218B116.20 GiB0.76 GiB117.90 GiB1.14 GiB65±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.06 GiB117.43 GiB1.61 GiB100±37%
step-3.5-flashQ4_1199B115.15 GiB1.34 GiB117.42 GiB1.62 GiB18±26.5%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.06 GiB116.84 GiB2.20 GiB111±37%
grok-2MoEIQ3_M270B115.25 GiB0.53 GiB116.82 GiB2.22 GiB33±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.09 GiB116.74 GiB2.30 GiB99±37%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB0.76 GiB115.74 GiB3.30 GiB82±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB0.76 GiB115.74 GiB3.30 GiB82±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB0.51 GiB115.40 GiB3.64 GiB97±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB0.76 GiB115.26 GiB3.78 GiB83±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB0.51 GiB115.18 GiB3.86 GiB97±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB0.51 GiB115.18 GiB3.86 GiB97±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB0.51 GiB114.93 GiB4.11 GiB98±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB0.40 GiB113.80 GiB5.24 GiB124±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.50 GiB113.46 GiB5.58 GiB114±37%
Hermes-4-405BIQ2_XS406B111.15 GiB1.05 GiB113.38 GiB5.66 GiB19±26.5%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB1.05 GiB113.38 GiB5.66 GiB19±26.5%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.06 GiB113.25 GiB5.79 GiB114±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB0.40 GiB112.96 GiB6.08 GiB105±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB0.51 GiB111.62 GiB7.42 GiB100±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB0.38 GiB110.70 GiB8.34 GiB80±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB0.38 GiB110.70 GiB8.34 GiB80±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.03 GiB110.24 GiB8.80 GiB113±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB0.76 GiB110.18 GiB8.86 GiB77±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB0.25 GiB109.77 GiB9.27 GiB103±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB0.51 GiB109.30 GiB9.74 GiB78±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.03 GiB109.08 GiB9.96 GiB114±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB0.46 GiB109.02 GiB10.02 GiB36±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB0.76 GiB108.85 GiB10.19 GiB87±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB0.40 GiB107.99 GiB11.05 GiB82±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.18 GiB107.95 GiB11.09 GiB96±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB0.51 GiB107.80 GiB11.24 GiB87±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB0.51 GiB107.80 GiB11.24 GiB87±37%
GLM-4.6VMoEQ8_0108B105.81 GiB0.38 GiB107.12 GiB11.92 GiB82±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.06 GiB106.41 GiB12.63 GiB102±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB0.95 GiB104.92 GiB14.12 GiB20±26.5%
dots.llm1.instMoEQ5_K_M143B101.84 GiB2.06 GiB104.83 GiB14.21 GiB84±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB0.53 GiB104.35 GiB14.69 GiB20±26.5%
gpt-oss-20b-hereticMoEQ8_020.9B102.72 GiB0.06 GiB103.67 GiB15.37 GiB61±37%
Huihui-gpt-oss-20b-BF16-abliteratedMoEQ8_020.9B102.72 GiB0.06 GiB103.67 GiB15.37 GiB61±37%
gemma-2-27b-itF3227.2B101.43 GiB0.76 GiB103.23 GiB15.81 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 q8_0 KV cache, the largest being Hy3 at IQ3_XXS. 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.