AMD · datacenter

Instinct MI210

Instinct MI210 has 64 GB of VRAM at 1638 GB/s — about 59.52 GiB usable after driver and compositor overhead. 2056 of 2118 indexed models fit at 32K context with q8_0 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
64 GB
HBM2e
Bandwidth
1638 GB/s
4096-bit bus
Tensor FP16
181 TF
dense
TDP
300 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1767vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2056 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3.5-88BMoEI1-Q5_K_M87.7B58.10 GiB0.40 GiB59.43 GiB0.09 GiB84±37%
Qwen3-Coder-30B-A3B-InstructMoEBF1630.5B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Salience-1.5-FlashMoEBF1631.1B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Qwen3-VL-30B-A3B-InstructMoEBF1631.1B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Qwen3-VL-30B-A3B-ThinkingMoEBF1631.1B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
MiroThinker-v1.0-30BMoEBF1630.5B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Qwen3-30B-A3BMoEBF1630.5B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Qwen3-30B-A3B-Thinking-2507-Claude-4.5-Sonnet-High-Reasoning-DistillMoEBF1630.5B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Qwen3-30B-A3B-Instruct-2507MoEBF1630.5B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Qwen3-30B-A3B-Thinking-2507MoEBF1630.5B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Pantheon-Proto-RP-1.8-30B-A3BMoEBF1630.5B56.90 GiB1.59 GiB59.39 GiB0.13 GiB66±37%
Tongyi-DeepResearch-30B-A3BMoEBF1630.5B56.90 GiB1.59 GiB59.38 GiB0.14 GiB66±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_XXS236B57.28 GiB1.12 GiB59.34 GiB0.18 GiB81±37%
DeepSeek-V2.5MoEIQ2_XXS236B57.28 GiB1.12 GiB59.34 GiB0.18 GiB81±37%
DeepSeek-Coder-V2-InstructMoEIQ2_XXS236B57.28 GiB1.12 GiB59.34 GiB0.18 GiB81±37%
Phi-3-mini-4k-instructKV unresolvedF323.8B52.01 GiB6.38 GiB59.29 GiB0.23 GiB18±26.5%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ3_K_S124B56.93 GiB1.46 GiB59.29 GiB0.23 GiB75±37%
step-3.5-flashIQ2_XS199B51.40 GiB6.92 GiB59.25 GiB0.27 GiB18±26.5%
Qwen3.5-99BMoEI1-Q4_199.0B57.88 GiB0.40 GiB59.20 GiB0.32 GiB88±37%
Qwen2.5-14B-Instruct-1MF3214.8B55.03 GiB3.19 GiB59.16 GiB0.36 GiB18±26.5%
DeepSeek-R1-Distill-Qwen-14BF3214.8B55.03 GiB3.19 GiB59.16 GiB0.36 GiB18±26.5%
Step-3.7-FlashIQ2_XXS201B51.22 GiB6.92 GiB59.07 GiB0.45 GiB18±26.5%
WizardLM-Uncensored-SuperCOT-StoryTelling-30bQ8_032.5B32.19 GiB25.90 GiB59.06 GiB0.46 GiB18±26.5%
archangel_sft-kto_llama30bQ8_032.5B32.19 GiB25.90 GiB59.06 GiB0.46 GiB18±26.5%
Wizard-Vicuna-30B-UncensoredQ8_032.5B32.19 GiB25.90 GiB59.06 GiB0.46 GiB18±26.5%
granite-4.1-30bBF1628.9B53.77 GiB4.25 GiB59.03 GiB0.49 GiB18±26.5%
Qwen3.5-122B-A10BMoEUD-IQ4_XS125B57.67 GiB0.40 GiB59.00 GiB0.52 GiB94±37%
Mixtral-8x22B-Instruct-v0.1MoEIQ3_XS141B54.23 GiB3.72 GiB58.91 GiB0.61 GiB30±37%
Mixtral-8x22B-v0.1MoEIQ3_XS141B54.23 GiB3.72 GiB58.91 GiB0.61 GiB30±37%
Mixtral-8x22B-v0.1MoEIQ3_XS141B54.23 GiB3.72 GiB58.91 GiB0.61 GiB30±37%
phi-4F3214.7B54.61 GiB3.32 GiB58.89 GiB0.63 GiB18±26.5%
Llama-3_1-Nemotron-51B-InstructIQ2_S51.5B15.33 GiB42.50 GiB58.87 GiB0.65 GiB18±26.5%
GLM-4.6VMoEIQ4_XS108B54.80 GiB3.05 GiB58.79 GiB0.73 GiB58±37%
Qwen3-14B-GPT-5.2-High-Reasoning-DistillBF1614.8B55.03 GiB2.66 GiB58.65 GiB0.87 GiB18±26.5%
GLM-4.7-REAP-218B-A32BMoEIQ2_XXS218B51.57 GiB6.11 GiB58.62 GiB0.90 GiB43±37%
Hy-MT2-30B-A3BMoEBF1630.1B56.03 GiB1.59 GiB58.52 GiB1.00 GiB66±37%
openPangu-2.0-FlashMoEKV unresolvedQ4_K_M100B56.71 GiB0.86 GiB58.48 GiB1.04 GiB89±37%
Behemoth-X-123B-v2IQ3_M123B51.48 GiB5.84 GiB58.38 GiB1.14 GiB18±26.5%
Mistral-Large-Instruct-2411IQ3_M123B51.48 GiB5.84 GiB58.38 GiB1.14 GiB18±26.5%
Llama-3_3-Nemotron-Super-49B-v1_5IQ2_S49.9B14.76 GiB42.50 GiB58.30 GiB1.22 GiB18±26.5%
Valkyrie-49B-v2.1I1-IQ2_S49.9B14.76 GiB42.50 GiB58.30 GiB1.22 GiB18±26.5%
Llama-3_3-Nemotron-Super-49B-v1IQ2_S49.9B14.76 GiB42.50 GiB58.30 GiB1.22 GiB18±26.5%
North-Mini-Code-1.0MoEBF1630.5B56.81 GiB0.60 GiB58.29 GiB1.23 GiB74±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEQ8_057.3B55.03 GiB2.22 GiB58.20 GiB1.32 GiB63±37%
GLM-4.5VMoEI1-IQ4_XS108B54.13 GiB3.05 GiB58.11 GiB1.41 GiB59±37%
Mistral-Medium-3.5-128BIQ3_XS128B51.16 GiB5.84 GiB58.06 GiB1.46 GiB18±26.5%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ3_XS139B53.00 GiB4.12 GiB58.00 GiB1.52 GiB56±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ3_XS139B53.00 GiB4.12 GiB58.00 GiB1.52 GiB56±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ3_K_L109B53.83 GiB3.19 GiB57.94 GiB1.58 GiB58±37%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ8_027.8B55.90 GiB1.06 GiB57.93 GiB1.59 GiB18±26.5%
Qwen3-Omni-30B-A3B-InstructBF1635.3B56.90 GiB0.00 GiB57.85 GiB1.67 GiB18±26.5%
Qwen3-Omni-30B-A3B-ThinkingBF1631.7B56.90 GiB0.00 GiB57.85 GiB1.67 GiB18±26.5%
InternVL3_5-30B-A3BBF1630.8B56.90 GiB0.00 GiB57.84 GiB1.68 GiB18±26.5%
Qwen3.8-27BBF1627.8B55.65 GiB1.06 GiB57.68 GiB1.84 GiB18±26.5%
c4ai-command-r-plus-08-2024IQ4_XS104B52.34 GiB4.25 GiB57.67 GiB1.85 GiB18±26.5%
XORTRON-NXTXPRTXXLI1-IQ3_S128B50.77 GiB5.84 GiB57.67 GiB1.85 GiB18±26.5%
Qwen3-Coder-NextMoEUD-Q5_K_M79.7B55.17 GiB1.59 GiB57.66 GiB1.86 GiB89±37%
GLM-4.7-FlashMoEBF1631.2B55.79 GiB0.88 GiB57.58 GiB1.94 GiB72±37%
GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-DistillMoEBF1631.2B55.79 GiB0.88 GiB57.58 GiB1.94 GiB72±37%
GLM-4.7-Flash-hereticMoEBF1629.9B55.79 GiB0.88 GiB57.58 GiB1.94 GiB72±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.

Questions people ask

What AI models can a Instinct MI210 run?
2056 of 2118 indexed open-weight models fit a Instinct MI210 at 32,768 context with q8_0 KV cache, the largest being Qwen3.5-88B at I1-Q5_K_M. That covers text, vision-language, image, video and speech models.
How much usable memory does a Instinct MI210 actually have?
Its nameplate is 64 GB, but about 59.52 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Instinct MI210 fast for local AI?
Its memory bandwidth is 1638 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.