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 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 8K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-235B-A22B-Instruct-2507MoEIQ4_XS235B117.24 GiB0.78 GiB118.95 GiB0.09 GiB74±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_XS235B117.24 GiB0.78 GiB118.95 GiB0.09 GiB74±37%
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.12 GiB118.88 GiB0.16 GiB109±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.09 GiB118.81 GiB0.23 GiB98±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_XXS311B117.27 GiB0.50 GiB118.72 GiB0.32 GiB95±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.36 GiB118.68 GiB0.36 GiB76±37%
GLM-4.7-REAP-218B-A32BMoEIQ4_NL218B116.20 GiB1.53 GiB118.66 GiB0.38 GiB63±37%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB0.78 GiB118.60 GiB0.44 GiB74±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_XS310B117.12 GiB0.50 GiB118.57 GiB0.47 GiB95±37%
ERNIE-4.5-300B-A47B-PTUD-IQ3_XXS300B116.56 GiB0.90 GiB118.48 GiB0.56 GiB18±26.5%
Qwen3-VL-235B-A22B-ThinkingMoEIQ4_XS236B116.70 GiB0.78 GiB118.41 GiB0.63 GiB74±37%
Qwen3-VL-235B-A22B-InstructMoEIQ4_XS236B116.70 GiB0.78 GiB118.41 GiB0.63 GiB74±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ4_XS235B116.68 GiB0.78 GiB118.39 GiB0.65 GiB74±37%
step-3.5-flashQ4_1199B115.15 GiB2.14 GiB118.22 GiB0.82 GiB18±26.5%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.28 GiB118.16 GiB0.88 GiB92±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.28 GiB118.16 GiB0.88 GiB92±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.28 GiB118.16 GiB0.88 GiB92±37%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.67 GiB118.10 GiB0.94 GiB109±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.67 GiB118.10 GiB0.94 GiB109±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.12 GiB117.49 GiB1.55 GiB100±37%
grok-2MoEIQ3_M270B115.25 GiB1.06 GiB117.35 GiB1.69 GiB33±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.12 GiB116.91 GiB2.13 GiB110±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.16 GiB116.82 GiB2.22 GiB99±37%
Step-3.7-FlashUD-Q4_K_M201B113.71 GiB2.14 GiB116.78 GiB2.26 GiB18±26.5%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB1.53 GiB116.50 GiB2.54 GiB78±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB1.53 GiB116.50 GiB2.54 GiB78±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB1.53 GiB116.02 GiB3.02 GiB79±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB1.03 GiB115.91 GiB3.13 GiB93±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB1.03 GiB115.69 GiB3.35 GiB93±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB1.03 GiB115.69 GiB3.35 GiB93±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB1.03 GiB115.44 GiB3.60 GiB94±37%
Hermes-4-405BIQ2_XS406B111.15 GiB2.09 GiB114.43 GiB4.61 GiB18±26.5%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB2.09 GiB114.43 GiB4.61 GiB18±26.5%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB0.80 GiB114.20 GiB4.84 GiB119±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.67 GiB113.63 GiB5.41 GiB112±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB0.80 GiB113.36 GiB5.68 GiB102±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.12 GiB113.31 GiB5.73 GiB113±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB1.03 GiB112.14 GiB6.90 GiB96±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB0.76 GiB111.08 GiB7.96 GiB78±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB0.76 GiB111.08 GiB7.96 GiB78±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB1.53 GiB110.94 GiB8.10 GiB73±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.03 GiB110.24 GiB8.80 GiB113±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB0.50 GiB110.02 GiB9.02 GiB101±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB1.03 GiB109.81 GiB9.23 GiB76±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB1.53 GiB109.61 GiB9.43 GiB82±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB0.93 GiB109.49 GiB9.55 GiB35±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.03 GiB109.08 GiB9.96 GiB114±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB0.80 GiB108.39 GiB10.65 GiB80±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB1.03 GiB108.32 GiB10.72 GiB84±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB1.03 GiB108.32 GiB10.72 GiB84±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.37 GiB108.13 GiB10.91 GiB94±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB10.63 GiB107.60 GiB11.44 GiB19±26.5%
GLM-4.6VMoEQ8_0108B105.81 GiB0.76 GiB107.50 GiB11.54 GiB80±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB4.12 GiB106.89 GiB12.15 GiB74±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.12 GiB106.47 GiB12.57 GiB102±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB1.29 GiB105.25 GiB13.79 GiB20±26.5%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB1.06 GiB104.88 GiB14.16 GiB20±26.5%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB10.63 GiB104.56 GiB14.48 GiB20±26.5%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB10.63 GiB104.56 GiB14.48 GiB20±26.5%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB10.63 GiB104.56 GiB14.48 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 q8_0 KV cache, the largest being Qwen3-235B-A22B-Instruct-2507 at IQ4_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.