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 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
vision language 191text 1796audio 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◐
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.26 GiB119.02 GiB0.02 GiB108±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.20 GiB118.92 GiB0.12 GiB97±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.40 GiB118.72 GiB0.32 GiB75±37%
grok-2MoEIQ3_M270B115.25 GiB2.25 GiB118.54 GiB0.50 GiB32±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.59 GiB118.47 GiB0.57 GiB90±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.59 GiB118.47 GiB0.57 GiB90±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.59 GiB118.47 GiB0.57 GiB90±37%
Step-3.7-FlashUD-Q4_K_M201B113.71 GiB3.67 GiB118.30 GiB0.74 GiB18±26.5%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB3.23 GiB118.21 GiB0.83 GiB71±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB3.23 GiB118.21 GiB0.83 GiB71±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.75 GiB118.18 GiB0.86 GiB108±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.75 GiB118.18 GiB0.86 GiB108±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB3.23 GiB117.73 GiB1.31 GiB71±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.26 GiB117.63 GiB1.41 GiB99±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB2.18 GiB117.06 GiB1.98 GiB86±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.26 GiB117.05 GiB1.99 GiB109±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.32 GiB116.98 GiB2.06 GiB97±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB2.18 GiB116.84 GiB2.20 GiB86±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB2.18 GiB116.84 GiB2.20 GiB86±37%
Hermes-4-405BIQ2_XS406B111.15 GiB4.43 GiB116.76 GiB2.28 GiB18±26.5%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB4.43 GiB116.76 GiB2.28 GiB18±26.5%
step-3.5-flashQ4_K_L199B112.07 GiB3.67 GiB116.66 GiB2.38 GiB18±26.5%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB2.18 GiB116.59 GiB2.45 GiB86±37%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB22.50 GiB116.44 GiB2.60 GiB18±26.5%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB22.50 GiB116.44 GiB2.60 GiB18±26.5%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB22.50 GiB116.44 GiB2.60 GiB18±26.5%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_L235B113.46 GiB1.65 GiB116.04 GiB3.00 GiB72±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_XXS310B113.14 GiB1.05 GiB115.14 GiB3.90 GiB94±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB1.69 GiB115.09 GiB3.95 GiB110±37%
GLM-4.7-REAP-218B-A32BMoEIQ4_XS218B110.09 GiB3.23 GiB114.27 GiB4.77 GiB60±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB1.69 GiB114.25 GiB4.79 GiB95±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.75 GiB113.71 GiB5.33 GiB111±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.26 GiB113.45 GiB5.59 GiB112±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB2.18 GiB113.29 GiB5.75 GiB88±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB3.23 GiB112.65 GiB6.39 GiB66±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB1.62 GiB111.93 GiB7.11 GiB74±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB1.62 GiB111.93 GiB7.11 GiB74±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB8.72 GiB111.49 GiB7.55 GiB58±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB3.23 GiB111.32 GiB7.72 GiB74±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB2.18 GiB110.96 GiB8.08 GiB71±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB1.05 GiB110.57 GiB8.47 GiB97±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB1.97 GiB110.52 GiB8.52 GiB34±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.02 GiB110.22 GiB8.82 GiB113±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB2.18 GiB109.47 GiB9.57 GiB78±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB2.18 GiB109.47 GiB9.57 GiB78±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB1.69 GiB109.28 GiB9.76 GiB75±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.02 GiB109.06 GiB9.98 GiB114±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB1.05 GiB108.98 GiB10.06 GiB98±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.77 GiB108.54 GiB10.50 GiB91±37%
GLM-4.6VMoEQ8_0108B105.81 GiB1.62 GiB108.35 GiB10.69 GiB76±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB1.65 GiB107.31 GiB11.73 GiB77±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB1.65 GiB107.31 GiB11.73 GiB77±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB1.65 GiB107.31 GiB11.73 GiB77±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB1.65 GiB107.31 GiB11.73 GiB77±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB1.65 GiB107.31 GiB11.73 GiB77±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.26 GiB106.61 GiB12.43 GiB100±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB2.25 GiB106.07 GiB12.97 GiB20±26.5%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB1.74 GiB105.70 GiB13.34 GiB20±26.5%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB3.09 GiB105.28 GiB13.76 GiB20±26.5%
Hy3MoEQ2_K299B101.28 GiB2.81 GiB105.03 GiB14.01 GiB83±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 q4_0 KV cache, the largest being Qwen3.5-397B-A17B at UD-IQ2_M. 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.