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 8K context with f16 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 8K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_XS310B117.12 GiB0.94 GiB119.01 GiB0.03 GiB92±37%
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.23 GiB118.99 GiB0.05 GiB108±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.67 GiB118.99 GiB0.05 GiB74±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.18 GiB118.89 GiB0.15 GiB97±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB1.26 GiB118.69 GiB0.35 GiB103±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB1.26 GiB118.69 GiB0.35 GiB103±37%
Step-3.7-FlashUD-Q4_K_M201B113.71 GiB4.03 GiB118.66 GiB0.38 GiB18±26.5%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.53 GiB118.41 GiB0.63 GiB90±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.53 GiB118.41 GiB0.63 GiB90±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.53 GiB118.41 GiB0.63 GiB90±37%
grok-2MoEIQ3_M270B115.25 GiB2.00 GiB118.29 GiB0.75 GiB32±37%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB2.88 GiB117.85 GiB1.19 GiB72±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB2.88 GiB117.85 GiB1.19 GiB72±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.23 GiB117.60 GiB1.44 GiB99±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB2.88 GiB117.37 GiB1.67 GiB72±37%
step-3.5-flashQ4_K_L199B112.07 GiB4.03 GiB117.03 GiB2.01 GiB18±26.5%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.23 GiB117.02 GiB2.02 GiB109±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB20.00 GiB116.98 GiB2.06 GiB18±26.5%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.31 GiB116.96 GiB2.08 GiB98±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB1.94 GiB116.82 GiB2.22 GiB87±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB1.94 GiB116.60 GiB2.44 GiB87±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB1.94 GiB116.60 GiB2.44 GiB87±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB1.94 GiB116.35 GiB2.69 GiB88±37%
Hermes-4-405BIQ2_XS406B111.15 GiB3.94 GiB116.27 GiB2.77 GiB18±26.5%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB3.94 GiB116.27 GiB2.77 GiB18±26.5%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_L235B113.46 GiB1.47 GiB115.86 GiB3.18 GiB73±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB1.50 GiB114.90 GiB4.14 GiB112±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB1.26 GiB114.22 GiB4.82 GiB106±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB1.50 GiB114.06 GiB4.98 GiB96±37%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB20.00 GiB113.94 GiB5.10 GiB18±26.5%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB20.00 GiB113.94 GiB5.10 GiB18±26.5%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB20.00 GiB113.94 GiB5.10 GiB18±26.5%
GLM-4.7-REAP-218B-A32BMoEIQ4_XS218B110.09 GiB2.88 GiB113.91 GiB5.13 GiB61±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.23 GiB113.42 GiB5.62 GiB112±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB1.94 GiB113.05 GiB5.99 GiB89±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB2.88 GiB112.29 GiB6.75 GiB68±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB1.44 GiB111.75 GiB7.29 GiB75±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB1.44 GiB111.75 GiB7.29 GiB75±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB2.88 GiB110.96 GiB8.08 GiB75±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB1.94 GiB110.72 GiB8.32 GiB72±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB7.75 GiB110.52 GiB8.52 GiB60±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB0.94 GiB110.46 GiB8.58 GiB98±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB1.75 GiB110.31 GiB8.73 GiB34±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.06 GiB110.26 GiB8.78 GiB112±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB1.94 GiB109.23 GiB9.81 GiB79±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB1.94 GiB109.23 GiB9.81 GiB79±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.06 GiB109.11 GiB9.93 GiB113±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB1.50 GiB109.09 GiB9.95 GiB76±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB0.94 GiB108.86 GiB10.18 GiB99±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.69 GiB108.45 GiB10.59 GiB92±37%
GLM-4.6VMoEQ8_0108B105.81 GiB1.44 GiB108.17 GiB10.87 GiB77±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB1.47 GiB107.13 GiB11.91 GiB77±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB1.47 GiB107.13 GiB11.91 GiB77±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB1.47 GiB107.13 GiB11.91 GiB77±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB1.47 GiB107.13 GiB11.91 GiB77±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB1.47 GiB107.13 GiB11.91 GiB77±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.23 GiB106.58 GiB12.46 GiB101±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB2.42 GiB106.38 GiB12.66 GiB20±26.5%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB2.00 GiB105.82 GiB13.22 GiB20±26.5%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB2.75 GiB104.93 GiB14.11 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 8,192 context with f16 KV cache, the largest being MiMo-V2-Flash at I1-IQ3_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.