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 128K 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 128K context

largest quantization that fits, per model · 2091 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
command-a-plus-05-2026-bf16MoEIQ4_NL219B116.79 GiB1.24 GiB118.94 GiB0.10 GiB72±37%
Hermes-4-405BIQ2_XXS406B99.91 GiB17.72 GiB118.81 GiB0.23 GiB18±26.5%
Hermes-3-Llama-3.1-405BIQ2_XXS406B99.91 GiB17.72 GiB118.81 GiB0.23 GiB18±26.5%
Qwen3.5-397B-A17BMoEIQ2_XS403B116.79 GiB1.05 GiB118.80 GiB0.24 GiB101±37%
step-3.5-flashIQ4_NL199B103.97 GiB13.79 GiB118.69 GiB0.35 GiB18±26.5%
dots.llm1.instMoEQ4_K_S143B82.81 GiB34.88 GiB118.62 GiB0.42 GiB27±37%
Llama-3_1-Nemotron-51B-InstructQ4_K_S51.5B27.46 GiB90.00 GiB118.50 GiB0.54 GiB18±26.5%
Qwen3.5-REAP-262B-A17BMoEQ3_K_M262B116.42 GiB1.05 GiB118.42 GiB0.62 GiB93±37%
MiMo-V2-FlashMoEKV unresolvedIQ3_XXS310B113.14 GiB4.22 GiB118.31 GiB0.73 GiB75±37%
gpt-oss-120b-abliteratedMoEQ8_0117B115.76 GiB1.27 GiB117.93 GiB1.11 GiB91±37%
Ornith-1.0-397BMoEUD-IQ2_M397B115.83 GiB1.05 GiB117.84 GiB1.20 GiB102±37%
Llama-3_3-Nemotron-Super-49B-v1_5Q4_K_S49.9B26.67 GiB90.00 GiB117.71 GiB1.33 GiB18±26.5%
Valkyrie-49B-v2.1I1-Q4_K_S49.9B26.67 GiB90.00 GiB117.71 GiB1.33 GiB18±26.5%
Llama-3_3-Nemotron-Super-49B-v1Q4_K_S49.9B26.67 GiB90.00 GiB117.71 GiB1.33 GiB18±26.5%
Qwen3-Coder-REAP-363B-A35BMoEUD-IQ1_M363B107.85 GiB8.72 GiB117.50 GiB1.54 GiB51±37%
DeepSeek-Coder-V2-Instruct-0724MoEQ3_K_L236B113.97 GiB2.37 GiB117.28 GiB1.76 GiB81±37%
DeepSeek-V2.5MoEQ3_K_L236B113.97 GiB2.37 GiB117.28 GiB1.76 GiB81±37%
DeepSeek-Coder-V2-InstructMoEQ3_K_L236B113.97 GiB2.37 GiB117.28 GiB1.76 GiB81±37%
GLM-4.5-AirMoEQ8_0110B109.39 GiB6.47 GiB116.79 GiB2.25 GiB57±37%
GLM-4.5-Air-DerestrictedMoEQ8_0110B109.39 GiB6.47 GiB116.79 GiB2.25 GiB57±37%
GLM-4.6-REAP-268B-A32BMoEUD-IQ3_XXS269B102.71 GiB12.94 GiB116.58 GiB2.46 GiB45±37%
Mixtral-8x22B-v0.1MoEQ6_K141B107.60 GiB7.88 GiB116.43 GiB2.61 GiB29±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-Q6_K139B106.40 GiB8.72 GiB116.01 GiB3.03 GiB55±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEQ6_K139B106.40 GiB8.72 GiB116.01 GiB3.03 GiB55±37%
Trinity-Large-ThinkingMoEIQ2_S399B112.03 GiB2.33 GiB115.29 GiB3.75 GiB97±37%
Step-3.7-FlashIQ4_XS201B99.94 GiB13.79 GiB114.65 GiB4.39 GiB18±26.5%
Mistral-Medium-3.5-128BQ6_K_L128B101.13 GiB12.38 GiB114.56 GiB4.48 GiB18±26.5%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ8_0109B106.67 GiB6.75 GiB114.34 GiB4.70 GiB57±37%
Nex-N2-ProMoEIQ2_S397B112.23 GiB1.05 GiB114.24 GiB4.80 GiB104±37%
GLM-4.5MoEUD-IQ1_M358B100.32 GiB12.94 GiB114.20 GiB4.84 GiB48±37%
grok-2MoEIQ3_XS270B104.12 GiB9.00 GiB114.17 GiB4.87 GiB29±37%
GLM-4.7MoEUD-IQ1_M358B100.27 GiB12.94 GiB114.14 GiB4.90 GiB48±37%
GLM-4.6MoEUD-IQ1_M357B100.02 GiB12.94 GiB113.90 GiB5.14 GiB48±37%
MiniMax-M3MoEIQ2_XXS427B108.60 GiB4.22 GiB113.74 GiB5.30 GiB77±37%
Hy3MoEQ2_K299B101.28 GiB11.25 GiB113.47 GiB5.57 GiB53±37%
GLM-4.6VMoEQ8_0108B105.81 GiB6.47 GiB113.21 GiB5.83 GiB58±37%
c4ai-command-r-plus-08-2024Q8_0104B102.74 GiB9.00 GiB112.82 GiB6.22 GiB19±26.5%
MiniMax-M2.7MoEUD-IQ4_NL229B103.15 GiB8.72 GiB112.75 GiB6.29 GiB62±37%
Qwen3-VL-235B-A22B-ThinkingMoEQ3_K_M236B104.72 GiB6.61 GiB112.27 GiB6.77 GiB58±37%
Qwen3-VL-235B-A22B-InstructMoEQ3_K_M236B104.72 GiB6.61 GiB112.27 GiB6.77 GiB58±37%
Qwen3-235B-A22BMoEQ3_K_M235B104.72 GiB6.61 GiB112.27 GiB6.77 GiB58±37%
Qwen3-235B-A22B-abliteratedMoEI1-Q3_K_M235B104.72 GiB6.61 GiB112.27 GiB6.77 GiB58±37%
Qwen3-235B-A22B-Thinking-2507MoEQ3_K_M235B104.72 GiB6.61 GiB112.27 GiB6.77 GiB58±37%
Qwen3-235B-A22B-Instruct-2507MoEQ3_K_M235B104.72 GiB6.61 GiB112.27 GiB6.77 GiB58±37%
MiMo-V2.5MoEKV unresolvedUD-IQ3_S311B106.98 GiB4.22 GiB112.14 GiB6.90 GiB78±37%
Trinity-Large-TrueBaseMoEI1-IQ2_S399B108.32 GiB2.33 GiB111.58 GiB7.46 GiB99±37%
GLM-4.7-REAP-218B-A32BMoEQ3_K_M218B97.57 GiB12.94 GiB111.45 GiB7.59 GiB43±37%
MiniMax-M2.1MoEQ3_K_M229B101.77 GiB8.72 GiB111.37 GiB7.67 GiB62±37%
MiniMax-M2.5MoEQ3_K_M229B101.77 GiB8.72 GiB111.37 GiB7.67 GiB62±37%
MiniMax-M2MoEQ3_K_M229B101.77 GiB8.72 GiB111.37 GiB7.67 GiB62±37%
MiniMax-M2.7-BF16-ultra-uncensored-hereticMoEQ3_K_M229B101.77 GiB8.72 GiB111.37 GiB7.67 GiB62±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q6_K124B106.87 GiB3.09 GiB110.86 GiB8.18 GiB78±37%
ERNIE-4.5-300B-A47B-PTQ2_K_L300B101.76 GiB7.59 GiB110.38 GiB8.66 GiB19±26.5%
DeepSeek-V4-FlashMoEUD-IQ3_S291B109.25 GiB0.02 GiB110.22 GiB8.82 GiB113±37%
Gemma-4-Dark-Gemistry-31BQ8_032.7B102.98 GiB5.95 GiB109.92 GiB9.12 GiB19±26.5%
gemma-2-27b-itF3227.2B101.43 GiB6.70 GiB109.16 GiB9.88 GiB19±26.5%
DeepSeek-V4-Flash-0731MoEUD-IQ3_S304B108.10 GiB0.02 GiB109.06 GiB9.98 GiB114±37%
Devstral-2-123B-Instruct-2512Q6_K125B95.53 GiB12.38 GiB108.96 GiB10.08 GiB19±26.5%
GLM-4.6-Derestricted-v3MoEIQ2_S357B94.86 GiB12.94 GiB108.74 GiB10.30 GiB49±37%
Qwen3.5-REAP-212B-A17BMoEIQ4_XS212B105.39 GiB1.05 GiB107.40 GiB11.64 GiB94±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 131,072 context with q4_0 KV cache, the largest being command-a-plus-05-2026-bf16 at IQ4_NL. 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.