What AI models can a Instinct MI210 run?

1940 of 2118 indexed open-weight models fit a Instinct MI210 at 131,072 context with f16 KV cache, the largest being Llama-4-Scout-17B-16E-Instruct at IQ2_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.

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. 1940 of 2118 indexed models fit at 128K 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

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

vision language 181text 1656image 1audio tts 21audio asr 39embedding 26video 16

What fits at 128K context

largest quantization that fits, per model · 1940 of 2118 indexed

Model	Best quant	Params	Weights●	KV●	Total in memory◐	Headroom◐	tok/s◐
Llama-4-Scout-17B-16E-InstructMoEKV unresolved	IQ2_M	109B	34.56 GiB	24.00 GiB	59.49 GiB	0.03 GiB	21±37%
Mixtral-8x22B-v0.1MoE	IQ1_M	141B	30.49 GiB	28.00 GiB	59.45 GiB	0.07 GiB	16±37%
Mixtral-8x22B-Instruct-v0.1MoE	IQ1_M	141B	30.49 GiB	28.00 GiB	59.45 GiB	0.07 GiB	16±37%
Mixtral-8x22B-v0.1MoE	IQ1_M	141B	30.48 GiB	28.00 GiB	59.44 GiB	0.08 GiB	16±37%
Laguna-S-2.1MoE	Q3_K_L	118B	52.33 GiB	6.14 GiB	59.39 GiB	0.13 GiB	50±37%
gemma-7b	I1-IQ2_XXS	8.5B	2.41 GiB	56.00 GiB	59.38 GiB	0.14 GiB	18±26.5%
MN-GRAND-23.5B-Gutenberg-UNCENSORED-V2-GLM4.7-Thinking	I1-Q6_K	23.4B	17.91 GiB	40.50 GiB	59.35 GiB	0.17 GiB	18±26.5%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolved	I1-Q2_K_S	109B	34.42 GiB	24.00 GiB	59.35 GiB	0.17 GiB	21±37%
Qwen3-Coder-REAP-25B-A3BMoE	BF16	24.9B	46.34 GiB	12.00 GiB	59.23 GiB	0.29 GiB	32±37%
Fallen-Gemma3-27B-v1	BF16	27.4B	50.32 GiB	7.94 GiB	59.20 GiB	0.32 GiB	18±26.5%
GLM-4.5-Air-DerestrictedMoE	IQ1_M	110B	35.26 GiB	23.00 GiB	59.19 GiB	0.33 GiB	22±37%
GLM-4.5-AirMoE	IQ1_M	110B	35.26 GiB	23.00 GiB	59.18 GiB	0.34 GiB	22±37%
DeepSeek-R1-Distill-Llama-70B	UD-IQ2_XXS	70.6B	18.11 GiB	40.00 GiB	59.14 GiB	0.38 GiB	18±26.5%
Hermes-4-70B	UD-IQ2_XXS	70.6B	18.10 GiB	40.00 GiB	59.12 GiB	0.40 GiB	18±26.5%
Llama-3.3-70B-Instruct	UD-IQ2_XXS	70.6B	18.10 GiB	40.00 GiB	59.12 GiB	0.40 GiB	18±26.5%
Assistant_Pepe_70B	IQ1_S	70.6B	18.07 GiB	40.00 GiB	59.09 GiB	0.43 GiB	18±26.5%
Qwen3.5-122B-A10BMoE	Q3_K_M	125B	55.15 GiB	3.00 GiB	59.08 GiB	0.44 GiB	69±37%
Qwen3.6-27B-uncensored-heretic-v2	BF16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
grug-27b	BF16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
Huihui-Qwen3.5-27B-abliterated	F16	27.8B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
Qwen3.6-27B-abliterated	F16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
Carnice-V2-27b	BF16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
Fara1.5-27B	BF16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
ThinkingCap-Qwen3.6-27B-heretic	F16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
Qwen-Image-Bench	BF16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
Qwen3.5-27B-uncensored-heretic-v1	BF16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-heretic	F16	27.4B	50.11 GiB	8.00 GiB	59.08 GiB	0.44 GiB	18±26.5%
Skyfall-31B-v4.2-heretic	Q8_0	31.4B	31.03 GiB	27.00 GiB	59.05 GiB	0.47 GiB	18±26.5%
Skyfall-31B-v4.2	Q8_0	31.4B	31.03 GiB	27.00 GiB	59.05 GiB	0.47 GiB	18±26.5%
GLM-4.6VMoE	IQ2_XXS	108B	35.11 GiB	23.00 GiB	59.03 GiB	0.49 GiB	22±37%
Qwopus3.5-122B-A10B-Kimi-K2.6-destill-healed-abliteratedMoE	IQ2_M	123B	55.04 GiB	3.00 GiB	58.97 GiB	0.55 GiB	69±37%
Apertus-70B-Instruct-2509	IQ2_XXS	70.6B	17.88 GiB	40.00 GiB	58.97 GiB	0.55 GiB	18±26.5%
IQuest-Coder-V1-40B-Instruct	I1-Q3_K_M	39.8B	17.92 GiB	40.00 GiB	58.91 GiB	0.61 GiB	18±26.5%
Mistral-Small-4-119B-2603MoE	UD-IQ4_NL	119B	55.13 GiB	2.81 GiB	58.88 GiB	0.64 GiB	70±37%
Maenad-70B	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-Reasoner	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Rombos-LLM-70b-Llama-3.3	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
L3.3-Electra-R1-70b	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
L3.3-70B-Magnum-v4-SE	IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Latxa-Llama-3.1-70B-Instruct-v2	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Llama-3.3_70_b_uncensored_continued	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Llama-3.3-70B-Instruct-abliterated	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
grok-oss-Revenant-70B	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Llama-3.1-Nemotron-70B-Instruct-HF	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
L3.3-70B-Euryale-v2.3	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Hermes-3-Llama-3.1-70B	IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Hermes-4-70B-heretic	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Llama-3.1-70B	IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Anubis-70B-v1.2	IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Golem-70B-v1b	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
DeepSeek-R1-Distill-Llama-70B-abliterated	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
DeepSeek-R1-Distill-Llama-70B-heretic	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Legion-V2.1-LLaMa-70B	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Tess-R1-Limerick-Llama-3.1-70B	IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
SEMIKONG-70B	IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
functionary-medium-v3.2KV unresolved	IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Llama-3.1-WhiteRabbitNeo-2-70B	IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Infinity-Instruct-7M-Gen-Llama3_1-70B	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
New-Dawn-Llama-3-70B-32K-v1.0	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 GiB	18±26.5%
Meta-Llama-3-70B-Instruct-abliterated-v3.5	I1-IQ2_XXS	70.6B	17.79 GiB	40.00 GiB	58.81 GiB	0.71 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?: 1940 of 2118 indexed open-weight models fit a Instinct MI210 at 131,072 context with f16 KV cache, the largest being Llama-4-Scout-17B-16E-Instruct at IQ2_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.