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. 2048 of 2118 indexed models fit at 128K context with q4_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 1759vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 128K context

largest quantization that fits, per model · 2048 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.5-Air-DerestrictedMoEQ3_K_L110B52.07 GiB6.47 GiB59.47 GiB0.05 GiB45±37%
GLM-4.5-AirMoEQ3_K_L110B52.07 GiB6.47 GiB59.47 GiB0.05 GiB45±37%
Apertus-70B-Instruct-2509Q5_K_M70.6B47.13 GiB11.25 GiB59.46 GiB0.06 GiB18±26.5%
step-3.5-flashIQ2_XXS199B44.74 GiB13.79 GiB59.45 GiB0.07 GiB18±26.5%
Qwen3.5-122B-A10BMoEUD-IQ4_XS125B57.67 GiB0.84 GiB59.45 GiB0.07 GiB88±37%
Qwen3-Coder-NextMoEUD-Q5_K_M79.7B55.17 GiB3.38 GiB59.44 GiB0.08 GiB71±37%
openPangu-2.0-FlashMoEKV unresolvedQ4_K_M100B56.71 GiB1.82 GiB59.44 GiB0.08 GiB78±37%
Llama-3.3-70B-Instruct-abliteratedQ5_K_L70.6B47.12 GiB11.25 GiB59.40 GiB0.12 GiB18±26.5%
Llama-3.3-70B-InstructQ5_K_L70.6B47.12 GiB11.25 GiB59.40 GiB0.12 GiB18±26.5%
Llama-3.1-Nemotron-70B-Instruct-HFQ5_K_L70.6B47.12 GiB11.25 GiB59.40 GiB0.12 GiB18±26.5%
Rombos-LLM-70b-Llama-3.3Q5_K_L70.6B47.12 GiB11.25 GiB59.40 GiB0.12 GiB18±26.5%
Anubis-70B-v1.2Q5_K_L70.6B47.12 GiB11.25 GiB59.40 GiB0.12 GiB18±26.5%
Tess-R1-Limerick-Llama-3.1-70BQ5_K_L70.6B47.12 GiB11.25 GiB59.40 GiB0.12 GiB18±26.5%
Infinity-Instruct-7M-Gen-Llama3_1-70BQ5_K_L70.6B47.12 GiB11.25 GiB59.40 GiB0.12 GiB18±26.5%
Athene-70BQ5_K_L70.6B47.12 GiB11.25 GiB59.40 GiB0.12 GiB18±26.5%
Mistral-Medium-3.5-128BUD-IQ3_XXS128B45.87 GiB12.38 GiB59.30 GiB0.22 GiB18±26.5%
Qwen2.5-72B-InstructQ5_072.7B46.88 GiB11.25 GiB59.16 GiB0.36 GiB18±26.5%
Qwen2.5-72BQ5_072.7B46.86 GiB11.25 GiB59.14 GiB0.38 GiB18±26.5%
Step-3.7-FlashIQ1_M201B44.41 GiB13.79 GiB59.12 GiB0.40 GiB18±26.5%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ8_027.8B55.90 GiB2.25 GiB59.12 GiB0.40 GiB18±26.5%
DeepSeek-V4-Flash-162BMoEKV unresolvedQ2_K92.2B55.09 GiB3.02 GiB59.06 GiB0.46 GiB65±37%
Hunyuan-A13B-InstructMoEQ5_K_M80.4B53.64 GiB4.50 GiB59.04 GiB0.48 GiB18±26.5%
GLM-4.6VMoEQ3_K_M108B51.64 GiB6.47 GiB59.03 GiB0.49 GiB45±37%
Devstral-2-123B-Instruct-2512UD-IQ3_XXS125B45.60 GiB12.38 GiB59.03 GiB0.49 GiB18±26.5%
command-r-35b-writer-v2I1-IQ3_XXS35.0B12.88 GiB45.00 GiB58.89 GiB0.63 GiB18±26.5%
GLM-4.5VMoEI1-Q3_K_M108B51.48 GiB6.47 GiB58.88 GiB0.64 GiB45±37%
Qwen3.8-27BBF1627.8B55.65 GiB2.25 GiB58.86 GiB0.66 GiB18±26.5%
Meta-Llama-3-70B-InstructQ5_K_M70.6B46.53 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Maenad-70BI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
calme-2.4-llama3-70bQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
calme-2.2-llama3-70bQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
L3.3-Electra-R1-70bI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
L3.3-70B-Magnum-v4-SEQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Llama-3.3_70_b_uncensored_continuedI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Strawberrylemonade-L3-70B-v1.2Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
grok-oss-Revenant-70BI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
L3.3-70B-Euryale-v2.3I1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Hermes-4-70B-hereticI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Hermes-4-70BQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Llama-3.1-70BQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Hermes-3-Llama-3.1-70BQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Golem-70B-v1bI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
DeepSeek-R1-Distill-Llama-70BQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
llama-3-firefunction-v2Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Legion-V2.1-LLaMa-70BI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Assistant_Pepe_70BI1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
SEMIKONG-70BQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
functionary-medium-v3.2KV unresolvedQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Llama-3.1-WhiteRabbitNeo-2-70BQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
New-Dawn-Llama-3-70B-32K-v1.0I1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-Q5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
L3.3-70B-Magnum-DiamondQ5_K_M70.6B46.52 GiB11.25 GiB58.80 GiB0.72 GiB18±26.5%
deepseek-llm-67b-chatI1-Q5_K_M67.4B44.38 GiB13.36 GiB58.74 GiB0.78 GiB18±26.5%
deepseek-llm-67b-baseI1-Q5_K_M67.4B44.38 GiB13.36 GiB58.74 GiB0.78 GiB18±26.5%
openbuddy-deepseek-67b-v15.3-4kI1-Q5_K_M67.4B44.38 GiB13.36 GiB58.74 GiB0.78 GiB18±26.5%
North-Mini-Code-1.0MoEBF1630.5B56.81 GiB1.00 GiB58.69 GiB0.83 GiB70±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?
2048 of 2118 indexed open-weight models fit a Instinct MI210 at 131,072 context with q4_0 KV cache, the largest being GLM-4.5-Air-Derestricted at Q3_K_L. 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.