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. 2050 of 2118 indexed models fit at 32K context with f16 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 1761vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 32K context

largest quantization that fits, per model · 2050 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Noromaid-20b-v0.1.1Q8_020.0B19.79 GiB38.75 GiB59.48 GiB0.04 GiB18±26.5%
Nethena-20BQ8_020.0B19.79 GiB38.75 GiB59.48 GiB0.04 GiB18±26.5%
Wizard-Vicuna-30B-UncensoredI1-IQ2_S32.5B9.67 GiB48.75 GiB59.39 GiB0.13 GiB18±26.5%
archangel_sft-kto_llama30bI1-IQ2_S32.5B9.67 GiB48.75 GiB59.39 GiB0.13 GiB18±26.5%
GLM-4.5-Air-REAP-82B-A12BMoEQ4_K_M81.9B52.68 GiB5.75 GiB59.36 GiB0.16 GiB44±37%
Qwen3.5-122B-A10BMoEUD-IQ4_XS125B57.67 GiB0.75 GiB59.35 GiB0.17 GiB89±37%
openPangu-2.0-FlashMoEKV unresolvedQ4_K_M100B56.71 GiB1.62 GiB59.23 GiB0.29 GiB80±37%
Assistant_Pepe_70BQ5_K_L70.6B48.20 GiB10.00 GiB59.23 GiB0.29 GiB18±26.5%
Phi-3-medium-128k-instructF3214.0B52.01 GiB6.25 GiB59.22 GiB0.30 GiB18±26.5%
Phi-3-medium-4k-instructF3214.0B52.01 GiB6.25 GiB59.22 GiB0.30 GiB18±26.5%
Delphi-25B-SimpleRL-MathQ8_025.0B24.71 GiB33.47 GiB59.16 GiB0.36 GiB18±26.5%
Mixtral-8x22B-v0.1MoEIQ3_XXS141B51.13 GiB7.00 GiB59.09 GiB0.43 GiB26±37%
Qwen2.5-Coder-32B-InstructQ6_K32.8B50.08 GiB8.00 GiB59.08 GiB0.44 GiB18±26.5%
Qwen3-Coder-NextMoEUD-Q5_K_M79.7B55.17 GiB3.00 GiB59.06 GiB0.46 GiB74±37%
command-r-35b-writer-v2IQ4_XS35.0B18.02 GiB40.00 GiB59.02 GiB0.50 GiB18±26.5%
GLM-4.7-REAP-218B-A32BMoEIQ1_M218B46.56 GiB11.50 GiB59.00 GiB0.52 GiB32±37%
CalmeRys-78B-Orpo-v0.1Q4_K_M78.0B47.22 GiB10.75 GiB59.00 GiB0.52 GiB18±26.5%
calme-2.3-rys-78bQ4_K_M78.0B47.22 GiB10.75 GiB59.00 GiB0.52 GiB18±26.5%
HuatuoGPT-o1-72BQ5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
Rombo-LLM-V3.0-Qwen-72bQ5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
EVA-Qwen2.5-72B-v0.2Q5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
MiroThinker-v1.0-72BQ5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
Qwen2.5-72BQ5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
magnum-v4-72bQ5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
KAT-Dev-72B-ExpQ5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
Homer-v1.0-Qwen2.5-72BQ5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
Chuluun-Qwen2.5-72B-v0.01Q5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
Qwen2.5-VL-72B-InstructQ5_K_S73.4B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q5_K_S72.7B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
UI-TARS-72B-DPOQ5_K_S73.4B47.85 GiB10.00 GiB58.88 GiB0.64 GiB18±26.5%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ8_027.8B55.90 GiB2.00 GiB58.87 GiB0.65 GiB18±26.5%
North-Mini-Code-1.0MoEBF1630.5B56.81 GiB1.13 GiB58.82 GiB0.70 GiB69±37%
Behemoth-X-123B-v2IQ3_XS123B46.70 GiB11.00 GiB58.75 GiB0.77 GiB18±26.5%
GLM-4.5-Air-DerestrictedMoEQ3_K_L110B52.07 GiB5.75 GiB58.75 GiB0.77 GiB47±37%
GLM-4.5-AirMoEQ3_K_L110B52.07 GiB5.75 GiB58.75 GiB0.77 GiB47±37%
DeepSeek-V4-Flash-162BMoEKV unresolvedQ2_K92.2B55.09 GiB2.69 GiB58.73 GiB0.79 GiB67±37%
step-3.5-flashIQ2_XXS199B44.74 GiB13.03 GiB58.70 GiB0.82 GiB18±26.5%
Qwen3.8-27BBF1627.8B55.65 GiB2.00 GiB58.61 GiB0.91 GiB18±26.5%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ3_XXS139B49.92 GiB7.75 GiB58.56 GiB0.96 GiB43±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ3_XXS139B49.92 GiB7.75 GiB58.56 GiB0.96 GiB43±37%
Hunyuan-A13B-InstructMoEQ5_K_M80.4B53.64 GiB4.00 GiB58.54 GiB0.98 GiB18±26.5%
Mistral-Medium-3.5-128BQ2_K128B46.44 GiB11.00 GiB58.49 GiB1.03 GiB18±26.5%
Step-3.7-FlashIQ1_M201B44.41 GiB13.03 GiB58.37 GiB1.15 GiB18±26.5%
GLM-4.7-FlashMoEBF1631.2B55.79 GiB1.65 GiB58.36 GiB1.16 GiB66±37%
GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-DistillMoEBF1631.2B55.79 GiB1.65 GiB58.36 GiB1.16 GiB66±37%
GLM-4.7-Flash-hereticMoEBF1629.9B55.79 GiB1.65 GiB58.36 GiB1.16 GiB66±37%
GLM-4.6VMoEQ3_K_M108B51.64 GiB5.75 GiB58.31 GiB1.21 GiB48±37%
gemma-2-27b-itBF1627.2B50.72 GiB6.56 GiB58.31 GiB1.21 GiB18±26.5%
magnum-v4-27bF1627.2B50.72 GiB6.56 GiB58.31 GiB1.21 GiB18±26.5%
Apertus-70B-Instruct-2509Q5_K_M70.6B47.13 GiB10.00 GiB58.21 GiB1.31 GiB18±26.5%
GLM-4.5VMoEI1-Q3_K_M108B51.48 GiB5.75 GiB58.16 GiB1.36 GiB48±37%
Llama-3.3-70B-Instruct-abliteratedQ5_K_L70.6B47.12 GiB10.00 GiB58.15 GiB1.37 GiB18±26.5%
Llama-3.3-70B-InstructQ5_K_L70.6B47.12 GiB10.00 GiB58.15 GiB1.37 GiB18±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFQ5_K_L70.6B47.12 GiB10.00 GiB58.15 GiB1.37 GiB18±26.5%
Rombos-LLM-70b-Llama-3.3Q5_K_L70.6B47.12 GiB10.00 GiB58.15 GiB1.37 GiB18±26.5%
Anubis-70B-v1.2Q5_K_L70.6B47.12 GiB10.00 GiB58.15 GiB1.37 GiB18±26.5%
Tess-R1-Limerick-Llama-3.1-70BQ5_K_L70.6B47.12 GiB10.00 GiB58.15 GiB1.37 GiB18±26.5%
Infinity-Instruct-7M-Gen-Llama3_1-70BQ5_K_L70.6B47.12 GiB10.00 GiB58.15 GiB1.37 GiB18±26.5%
Athene-70BQ5_K_L70.6B47.12 GiB10.00 GiB58.15 GiB1.37 GiB18±26.5%
v6-Finch-14B-HFF1614.1B26.63 GiB30.50 GiB58.08 GiB1.44 GiB18±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.

Questions people ask

What AI models can a Instinct MI210 run?
2050 of 2118 indexed open-weight models fit a Instinct MI210 at 32,768 context with f16 KV cache, the largest being Noromaid-20b-v0.1.1 at Q8_0. 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.