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 16K 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
vision language 191text 1796audio tts 21image 2audio asr 39video 16embedding 26

What fits at 16K 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 GiB0.25 GiB119.01 GiB0.03 GiB108±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.19 GiB118.91 GiB0.13 GiB97±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.49 GiB118.81 GiB0.23 GiB75±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.56 GiB118.44 GiB0.60 GiB90±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.56 GiB118.44 GiB0.60 GiB90±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.56 GiB118.44 GiB0.60 GiB90±37%
grok-2MoEIQ3_M270B115.25 GiB2.13 GiB118.42 GiB0.62 GiB32±37%
Step-3.7-FlashUD-Q4_K_M201B113.71 GiB3.74 GiB118.37 GiB0.67 GiB18±26.5%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.92 GiB118.35 GiB0.69 GiB106±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.92 GiB118.35 GiB0.69 GiB106±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB21.25 GiB118.23 GiB0.81 GiB18±26.5%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB3.05 GiB118.03 GiB1.01 GiB71±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB3.05 GiB118.03 GiB1.01 GiB71±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.25 GiB117.62 GiB1.42 GiB99±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB3.05 GiB117.55 GiB1.49 GiB71±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.25 GiB117.03 GiB2.01 GiB109±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.31 GiB116.97 GiB2.07 GiB98±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB2.06 GiB116.94 GiB2.10 GiB87±37%
step-3.5-flashQ4_K_L199B112.07 GiB3.74 GiB116.73 GiB2.31 GiB18±26.5%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB2.06 GiB116.72 GiB2.32 GiB87±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB2.06 GiB116.72 GiB2.32 GiB87±37%
Hermes-4-405BIQ2_XS406B111.15 GiB4.18 GiB116.52 GiB2.52 GiB18±26.5%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB4.18 GiB116.52 GiB2.52 GiB18±26.5%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB2.06 GiB116.47 GiB2.57 GiB87±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_L235B113.46 GiB1.56 GiB115.95 GiB3.09 GiB72±37%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB21.25 GiB115.19 GiB3.85 GiB18±26.5%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB21.25 GiB115.19 GiB3.85 GiB18±26.5%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB21.25 GiB115.19 GiB3.85 GiB18±26.5%
MiMo-V2-FlashMoEKV unresolvedIQ3_XXS310B113.14 GiB1.00 GiB115.08 GiB3.96 GiB94±37%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB1.59 GiB115.00 GiB4.04 GiB111±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB1.59 GiB114.15 GiB4.89 GiB96±37%
GLM-4.7-REAP-218B-A32BMoEIQ4_XS218B110.09 GiB3.05 GiB114.09 GiB4.95 GiB61±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.92 GiB113.88 GiB5.16 GiB110±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.25 GiB113.43 GiB5.61 GiB112±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB2.06 GiB113.17 GiB5.87 GiB89±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB3.05 GiB112.47 GiB6.57 GiB67±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB1.53 GiB111.84 GiB7.20 GiB75±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB1.53 GiB111.84 GiB7.20 GiB75±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB3.05 GiB111.14 GiB7.90 GiB74±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB8.23 GiB111.01 GiB8.03 GiB59±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB2.06 GiB110.84 GiB8.20 GiB71±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB1.00 GiB110.51 GiB8.53 GiB97±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB1.86 GiB110.42 GiB8.62 GiB34±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.03 GiB110.24 GiB8.80 GiB113±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB2.06 GiB109.35 GiB9.69 GiB79±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB2.06 GiB109.35 GiB9.69 GiB79±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB1.59 GiB109.19 GiB9.85 GiB76±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.03 GiB109.08 GiB9.96 GiB114±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB1.00 GiB108.92 GiB10.12 GiB98±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.73 GiB108.50 GiB10.54 GiB92±37%
GLM-4.6VMoEQ8_0108B105.81 GiB1.53 GiB108.26 GiB10.78 GiB76±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB1.56 GiB107.22 GiB11.82 GiB77±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB1.56 GiB107.22 GiB11.82 GiB77±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB1.56 GiB107.22 GiB11.82 GiB77±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB1.56 GiB107.22 GiB11.82 GiB77±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB1.56 GiB107.22 GiB11.82 GiB77±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.25 GiB106.59 GiB12.45 GiB101±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB2.13 GiB105.94 GiB13.10 GiB20±26.5%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB1.95 GiB105.91 GiB13.13 GiB20±26.5%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB2.92 GiB105.11 GiB13.93 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 16,384 context with q8_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.