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 64K 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 1761vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2050 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
HuatuoGPT-o1-72BQ5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
Rombo-LLM-V3.0-Qwen-72bQ5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
EVA-Qwen2.5-72B-v0.2Q5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
MiroThinker-v1.0-72BQ5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
Qwen2.5-72BQ5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
magnum-v4-72bQ5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
KAT-Dev-72B-ExpQ5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
Homer-v1.0-Qwen2.5-72BQ5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
Chuluun-Qwen2.5-72B-v0.01Q5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
Qwen2.5-VL-72B-InstructQ5_K_S73.4B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
Tower-Plus-72B-ultra-uncensored-hereticI1-Q5_K_S72.7B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
UI-TARS-72B-DPOQ5_K_S73.4B47.85 GiB10.63 GiB59.50 GiB0.02 GiB18±26.5%
Behemoth-X-123B-v2IQ3_XS123B46.70 GiB11.69 GiB59.44 GiB0.08 GiB18±26.5%
Qwen3.5-122B-A10BMoEUD-IQ4_XS125B57.67 GiB0.80 GiB59.40 GiB0.12 GiB89±37%
Wizard-Vicuna-30B-UncensoredI1-IQ1_S32.5B6.63 GiB51.80 GiB59.40 GiB0.12 GiB18±26.5%
archangel_sft-kto_llama30bI1-IQ1_S32.5B6.63 GiB51.80 GiB59.40 GiB0.12 GiB18±26.5%
openPangu-2.0-FlashMoEKV unresolvedQ4_K_M100B56.71 GiB1.72 GiB59.33 GiB0.19 GiB79±37%
Qwen3-Coder-NextMoEUD-Q5_K_M79.7B55.17 GiB3.19 GiB59.25 GiB0.27 GiB72±37%
Assistant_Pepe_70BQ5_K_M70.6B47.60 GiB10.63 GiB59.25 GiB0.27 GiB18±26.5%
Mistral-Medium-3.5-128BQ2_K128B46.44 GiB11.69 GiB59.18 GiB0.34 GiB18±26.5%
GLM-4.5-Air-DerestrictedMoEQ3_K_L110B52.07 GiB6.11 GiB59.11 GiB0.41 GiB46±37%
GLM-4.5-AirMoEQ3_K_L110B52.07 GiB6.11 GiB59.11 GiB0.41 GiB46±37%
command-r-35b-writer-v2I1-IQ3_M35.0B15.55 GiB42.50 GiB59.06 GiB0.46 GiB18±26.5%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ3_XXS139B49.92 GiB8.23 GiB59.04 GiB0.48 GiB42±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ3_XXS139B49.92 GiB8.23 GiB59.04 GiB0.48 GiB42±37%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ8_027.8B55.90 GiB2.13 GiB58.99 GiB0.53 GiB18±26.5%
step-3.5-flashIQ2_XXS199B44.74 GiB13.30 GiB58.96 GiB0.56 GiB18±26.5%
DeepSeek-V4-Flash-162BMoEKV unresolvedQ2_K92.2B55.09 GiB2.86 GiB58.89 GiB0.63 GiB66±37%
Apertus-70B-Instruct-2509Q5_K_M70.6B47.13 GiB10.63 GiB58.84 GiB0.68 GiB18±26.5%
Hunyuan-A13B-InstructMoEQ5_K_M80.4B53.64 GiB4.25 GiB58.79 GiB0.73 GiB18±26.5%
Llama-3.3-70B-Instruct-abliteratedQ5_K_L70.6B47.12 GiB10.63 GiB58.77 GiB0.75 GiB18±26.5%
Llama-3.3-70B-InstructQ5_K_L70.6B47.12 GiB10.63 GiB58.77 GiB0.75 GiB18±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFQ5_K_L70.6B47.12 GiB10.63 GiB58.77 GiB0.75 GiB18±26.5%
Rombos-LLM-70b-Llama-3.3Q5_K_L70.6B47.12 GiB10.63 GiB58.77 GiB0.75 GiB18±26.5%
Anubis-70B-v1.2Q5_K_L70.6B47.12 GiB10.63 GiB58.77 GiB0.75 GiB18±26.5%
Tess-R1-Limerick-Llama-3.1-70BQ5_K_L70.6B47.12 GiB10.63 GiB58.77 GiB0.75 GiB18±26.5%
Infinity-Instruct-7M-Gen-Llama3_1-70BQ5_K_L70.6B47.12 GiB10.63 GiB58.77 GiB0.75 GiB18±26.5%
Athene-70BQ5_K_L70.6B47.12 GiB10.63 GiB58.77 GiB0.75 GiB18±26.5%
Qwen3.8-27BBF1627.8B55.65 GiB2.13 GiB58.74 GiB0.78 GiB18±26.5%
North-Mini-Code-1.0MoEBF1630.5B56.81 GiB1.03 GiB58.72 GiB0.80 GiB70±37%
GLM-4.6VMoEQ3_K_M108B51.64 GiB6.11 GiB58.67 GiB0.85 GiB46±37%
Step-3.7-FlashIQ1_M201B44.41 GiB13.30 GiB58.63 GiB0.89 GiB18±26.5%
Qwen2.5-72B-InstructQ5_072.7B46.88 GiB10.63 GiB58.53 GiB0.99 GiB18±26.5%
GLM-4.5VMoEI1-Q3_K_M108B51.48 GiB6.11 GiB58.52 GiB1.00 GiB46±37%
GLM-4.7-FlashMoEBF1631.2B55.79 GiB1.76 GiB58.46 GiB1.06 GiB65±37%
GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-DistillMoEBF1631.2B55.79 GiB1.76 GiB58.46 GiB1.06 GiB65±37%
GLM-4.7-Flash-hereticMoEBF1629.9B55.79 GiB1.76 GiB58.46 GiB1.06 GiB65±37%
Devstral-2-123B-Instruct-2512UD-IQ3_XXS125B45.60 GiB11.69 GiB58.35 GiB1.17 GiB18±26.5%
gemma-2-27b-itBF1627.2B50.72 GiB6.54 GiB58.29 GiB1.23 GiB18±26.5%
magnum-v4-27bF1627.2B50.72 GiB6.54 GiB58.29 GiB1.23 GiB18±26.5%
Meta-Llama-3-70B-InstructQ5_K_M70.6B46.53 GiB10.63 GiB58.18 GiB1.34 GiB18±26.5%
Maenad-70BI1-Q5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 GiB18±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 GiB18±26.5%
calme-2.4-llama3-70bQ5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 GiB18±26.5%
calme-2.2-llama3-70bQ5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 GiB18±26.5%
L3.3-Electra-R1-70bI1-Q5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 GiB18±26.5%
L3.3-70B-Magnum-v4-SEQ5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 GiB18±26.5%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 GiB18±26.5%
Llama-3.3_70_b_uncensored_continuedI1-Q5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 GiB18±26.5%
Strawberrylemonade-L3-70B-v1.2Q5_K_M70.6B46.52 GiB10.63 GiB58.17 GiB1.35 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 65,536 context with q8_0 KV cache, the largest being HuatuoGPT-o1-72B at Q5_K_S. 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.
Instinct MI210 — what AI models can it run locally? — ossmodeldb