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 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
text 1796vision language 191audio 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◐
Qwen3-235B-A22B-Instruct-2507MoEIQ4_XS235B117.24 GiB0.83 GiB119.00 GiB0.04 GiB73±37%
Qwen3-235B-A22B-Thinking-2507MoEIQ4_XS235B117.24 GiB0.83 GiB119.00 GiB0.04 GiB73±37%
Qwen3.5-397B-A17BMoEUD-IQ2_M403B117.81 GiB0.13 GiB118.89 GiB0.15 GiB109±37%
Mistral-Small-4-119B-2603MoEQ8_0119B117.79 GiB0.10 GiB118.82 GiB0.22 GiB98±37%
GLM-4.7-REAP-218B-A32BMoEIQ4_NL218B116.20 GiB1.62 GiB118.75 GiB0.29 GiB62±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_XXS311B117.27 GiB0.53 GiB118.75 GiB0.29 GiB95±37%
Qwen3-235B-A22BMoEIQ4_XS235B116.89 GiB0.83 GiB118.64 GiB0.40 GiB74±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ3_XS310B117.12 GiB0.53 GiB118.60 GiB0.44 GiB95±37%
command-a-plus-05-2026-bf16MoEQ4_0219B117.42 GiB0.26 GiB118.58 GiB0.46 GiB76±37%
ERNIE-4.5-300B-A47B-PTUD-IQ3_XXS300B116.56 GiB0.95 GiB118.53 GiB0.51 GiB18±26.5%
Qwen3-VL-235B-A22B-ThinkingMoEIQ4_XS236B116.70 GiB0.83 GiB118.46 GiB0.58 GiB74±37%
Qwen3-VL-235B-A22B-InstructMoEIQ4_XS236B116.70 GiB0.83 GiB118.46 GiB0.58 GiB74±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ4_XS235B116.68 GiB0.83 GiB118.44 GiB0.60 GiB74±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ4_XS236B116.94 GiB0.30 GiB118.18 GiB0.86 GiB92±37%
DeepSeek-V2.5MoEIQ4_XS236B116.94 GiB0.30 GiB118.18 GiB0.86 GiB92±37%
DeepSeek-Coder-V2-InstructMoEIQ4_XS236B116.94 GiB0.30 GiB118.18 GiB0.86 GiB92±37%
step-3.5-flashQ4_1199B115.15 GiB1.98 GiB118.06 GiB0.98 GiB18±26.5%
Trinity-Large-PreviewMoEIQ2_M399B116.50 GiB0.49 GiB117.92 GiB1.12 GiB111±37%
Trinity-Large-TrueBaseMoEIQ2_M399B116.50 GiB0.49 GiB117.92 GiB1.12 GiB111±37%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB0.13 GiB117.50 GiB1.54 GiB100±37%
grok-2MoEIQ3_M270B115.25 GiB1.13 GiB117.42 GiB1.62 GiB33±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB0.13 GiB116.91 GiB2.13 GiB110±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB0.17 GiB116.82 GiB2.22 GiB99±37%
Step-3.7-FlashUD-Q4_K_M201B113.71 GiB1.98 GiB116.61 GiB2.43 GiB18±26.5%
GLM-4.5MoEUD-IQ2_M358B114.03 GiB1.62 GiB116.59 GiB2.45 GiB78±37%
GLM-4.7MoEUD-IQ2_M358B114.03 GiB1.62 GiB116.59 GiB2.45 GiB78±37%
GLM-4.6MoEUD-IQ2_M357B113.56 GiB1.62 GiB116.11 GiB2.93 GiB78±37%
MiniMax-M2.7MoEIQ4_XS229B114.00 GiB1.09 GiB115.97 GiB3.07 GiB93±37%
MiniMax-M2.1MoEIQ4_XS229B113.78 GiB1.09 GiB115.75 GiB3.29 GiB93±37%
MiniMax-M2MoEIQ4_XS229B113.78 GiB1.09 GiB115.75 GiB3.29 GiB93±37%
MiniMax-M2.5MoEIQ4_XS229B113.53 GiB1.09 GiB115.50 GiB3.54 GiB93±37%
Hermes-4-405BIQ2_XS406B111.15 GiB2.21 GiB114.55 GiB4.49 GiB18±26.5%
Hermes-3-Llama-3.1-405BIQ2_XS406B111.15 GiB2.21 GiB114.55 GiB4.49 GiB18±26.5%
Llama-4-Maverick-17B-128E-InstructMoEKV unresolvedUD-IQ1_S402B112.48 GiB0.84 GiB114.25 GiB4.79 GiB119±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB0.49 GiB113.44 GiB5.60 GiB114±37%
Solar-Open2-250BMoEQ3_K_M250B111.63 GiB0.84 GiB113.40 GiB5.64 GiB102±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB0.13 GiB113.32 GiB5.72 GiB113±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_L229B110.22 GiB1.09 GiB112.20 GiB6.84 GiB95±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB0.81 GiB111.13 GiB7.91 GiB78±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB0.81 GiB111.13 GiB7.91 GiB78±37%
GLM-4.6-REAP-268B-A32BMoEQ3_K_S269B108.47 GiB1.62 GiB111.03 GiB8.01 GiB73±37%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.02 GiB110.22 GiB8.82 GiB113±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB0.53 GiB110.05 GiB8.99 GiB101±37%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB1.09 GiB109.87 GiB9.17 GiB75±37%
GLM-4.6-Derestricted-v3MoEIQ2_M357B107.14 GiB1.62 GiB109.70 GiB9.34 GiB82±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB0.98 GiB109.54 GiB9.50 GiB35±37%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.02 GiB109.06 GiB9.98 GiB114±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB0.84 GiB108.44 GiB10.60 GiB79±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB1.09 GiB108.38 GiB10.66 GiB84±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB1.09 GiB108.38 GiB10.66 GiB84±37%
Llama-3_1-Nemotron-51B-InstructF1651.5B95.94 GiB11.25 GiB108.23 GiB10.81 GiB19±26.5%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB0.39 GiB108.15 GiB10.89 GiB94±37%
GLM-4.6VMoEQ8_0108B105.81 GiB0.81 GiB107.55 GiB11.49 GiB80±37%
dots.llm1.instMoEQ5_K_M143B101.84 GiB4.36 GiB107.13 GiB11.91 GiB73±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB0.13 GiB106.48 GiB12.56 GiB102±37%
Llama-3_3-Nemotron-Super-49B-v1_5BF1649.9B92.89 GiB11.25 GiB105.19 GiB13.85 GiB20±26.5%
Valkyrie-49B-v2.1BF1649.9B92.89 GiB11.25 GiB105.19 GiB13.85 GiB20±26.5%
Llama-3_3-Nemotron-Super-49B-v1BF1649.9B92.89 GiB11.25 GiB105.19 GiB13.85 GiB20±26.5%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB1.03 GiB104.99 GiB14.05 GiB20±26.5%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB1.13 GiB104.94 GiB14.10 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 q4_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.