What AI models can a Instinct MI210 run?

2027 of 2118 indexed open-weight models fit a Instinct MI210 at 131,072 context with q8_0 KV cache, the largest being MythoMax-L2-Kimiko-v2-13b at IQ3_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.

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. 2027 of 2118 indexed models fit at 128K 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

Context4K 8K 16K 32K 64K 128K

KV cachef16 q8_0 q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context

text 1739vision language 184image 2audio asr 39audio tts 21video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2027 of 2118 indexed

Model	Best quant	Params	Weights●	KV●	Total in memory◐	Headroom◐	tok/s◐
MythoMax-L2-Kimiko-v2-13b	IQ3_S	13.0B	5.45 GiB	53.13 GiB	59.52 GiB	0.00 GiB	18±26.5%
MythoMax-L2-13b	I1-IQ3_S	13.0B	5.45 GiB	53.13 GiB	59.52 GiB	0.00 GiB	18±26.5%
Meta-Llama-3-70B-Instruct	Q4_0	70.6B	37.23 GiB	21.25 GiB	59.51 GiB	0.01 GiB	18±26.5%
Huihui-GLM-4.5-Air-abliterated-lossytensorsMoE	I1-IQ3_XS	110B	46.34 GiB	12.22 GiB	59.48 GiB	0.04 GiB	33±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoE	UD-IQ3_S	124B	52.74 GiB	5.84 GiB	59.48 GiB	0.04 GiB	51±37%
OLMo-2-1124-13B-Instruct	IQ3_XS	13.7B	5.40 GiB	53.13 GiB	59.47 GiB	0.05 GiB	18±26.5%
GLM-4.6VMoE	UD-IQ3_XXS	108B	46.26 GiB	12.22 GiB	59.41 GiB	0.11 GiB	33±37%
deepseek-llm-67b-chat	I1-Q3_K_L	67.4B	33.13 GiB	25.23 GiB	59.37 GiB	0.15 GiB	18±26.5%
deepseek-llm-67b-base	I1-Q3_K_L	67.4B	33.13 GiB	25.23 GiB	59.37 GiB	0.15 GiB	18±26.5%
openbuddy-deepseek-67b-v15.3-4k	I1-Q3_K_L	67.4B	33.13 GiB	25.23 GiB	59.37 GiB	0.15 GiB	18±26.5%
Mixtral-8x22B-v0.1MoE	IQ2_M	141B	43.50 GiB	14.88 GiB	59.34 GiB	0.18 GiB	21±37%
codellama-13b-oasst-sft-v10	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
chronos-hermes-13b-v2	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
WhiteRabbitNeo-13B-v1	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
CodeLlama-13b-Instruct-hf	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
Orca-2-13b-Alpaca-Uncensored	I1-IQ3_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
WizardLM-13B-Uncensored	I1-IQ3_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
WizardCoder-Python-13B-V1.0	I1-IQ3_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
Guanaco-13B-Uncensored	I1-IQ3_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
Llama-2-13b-chat-hf	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
Wizard-Vicuna-13B-Uncensored	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
WizardLM-13b-V1.0-Uncensored	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
WizardLM-1.0-Uncensored-Llama2-13b	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
speechless-llama2-hermes-orca-platypus-wizardlm-13b	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
mythalion-13b	Q3_K_S	13.0B	5.27 GiB	53.13 GiB	59.34 GiB	0.18 GiB	18±26.5%
Kimi-Dev-72B	IQ4_XS	72.7B	37.02 GiB	21.25 GiB	59.30 GiB	0.22 GiB	18±26.5%
Qwen2.5-VL-72B-Instruct	IQ4_XS	73.4B	37.02 GiB	21.25 GiB	59.30 GiB	0.22 GiB	18±26.5%
Rombo-LLM-V3.0-Qwen-72b	I1-IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Qwen2.5-72B-Instruct-abliterated	I1-IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Qwen2.5-72B-Instruct-abliterated-v2	I1-IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
HuatuoGPT-o1-72B	IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
MiroThinker-v1.0-72B	I1-IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
EVA-Qwen2.5-72B-v0.2	IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Qwen2.5-Math-72B-Instruct	IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Qwen2.5-72B-Instruct	IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Malaysian-Qwen2.5-72B-Instruct	I1-IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Qwen2.5-72B	I1-IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
magnum-v4-72b	I1-IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
KAT-Dev-72B-Exp	IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Homer-v1.0-Qwen2.5-72B	IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Tower-Plus-72B-ultra-uncensored-heretic	I1-IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Chronos-Platinum-72B	IQ4_XS	72.7B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
UI-TARS-72B-DPO	IQ4_XS	73.4B	36.98 GiB	21.25 GiB	59.26 GiB	0.26 GiB	18±26.5%
Qwen3-Coder-NextMoE	UD-Q5_K_S	79.7B	51.99 GiB	6.38 GiB	59.25 GiB	0.27 GiB	54±37%
NSFW_13B_sft	Q2_K	13.3B	5.18 GiB	53.13 GiB	59.25 GiB	0.27 GiB	18±26.5%
Apertus-70B-Instruct-2509	Q3_K_L	70.6B	36.87 GiB	21.25 GiB	59.20 GiB	0.32 GiB	18±26.5%
CalmeRys-78B-Orpo-v0.1	I1-IQ3_M	78.0B	35.33 GiB	22.84 GiB	59.20 GiB	0.32 GiB	18±26.5%
calme-2.3-rys-78b	IQ3_M	78.0B	35.33 GiB	22.84 GiB	59.20 GiB	0.32 GiB	18±26.5%
Devstral-2-123B-Instruct-2512	IQ2_XS	125B	34.75 GiB	23.38 GiB	59.18 GiB	0.34 GiB	18±26.5%
Mistral-Medium-3.5-128B	I1-IQ2_XS	128B	34.75 GiB	23.38 GiB	59.18 GiB	0.34 GiB	18±26.5%
XORTRON-NXTXPRTXXL	I1-IQ2_XS	128B	34.75 GiB	23.38 GiB	59.18 GiB	0.34 GiB	18±26.5%
gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking	Q6_K	31.3B	46.94 GiB	11.25 GiB	59.17 GiB	0.35 GiB	18±26.5%
Chuluun-Qwen2.5-72B-v0.01	Q3_K_L	72.7B	36.79 GiB	21.25 GiB	59.07 GiB	0.45 GiB	18±26.5%
Qwen3-72B-Synthesis	Q3_K_L	72.7B	36.79 GiB	21.25 GiB	59.07 GiB	0.45 GiB	18±26.5%
GLM-4.5-Air-REAP-82B-A12BMoE	Q4_0	81.9B	45.78 GiB	12.22 GiB	58.93 GiB	0.59 GiB	31±37%
Laguna-S-2.1MoE	UD-IQ4_NL	118B	54.71 GiB	3.26 GiB	58.90 GiB	0.62 GiB	64±37%
Qwen3.5-88BMoE	I1-Q5_K_S	87.7B	56.33 GiB	1.59 GiB	58.85 GiB	0.67 GiB	74±37%
CodeLlama-70b-Instruct-hf	I1-Q4_K_S	69.0B	36.55 GiB	21.25 GiB	58.83 GiB	0.69 GiB	18±26.5%
CodeLlama-70b-Python-hf	I1-Q4_K_S	69.0B	36.55 GiB	21.25 GiB	58.83 GiB	0.69 GiB	18±26.5%
Nous-Hermes-Llama2-70b	I1-Q4_K_S	69.0B	36.55 GiB	21.25 GiB	58.83 GiB	0.69 GiB	18±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?: 2027 of 2118 indexed open-weight models fit a Instinct MI210 at 131,072 context with q8_0 KV cache, the largest being MythoMax-L2-Kimiko-v2-13b at IQ3_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.