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. 2063 of 2118 indexed models fit at 4K 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 1774vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 4K context

largest quantization that fits, per model · 2063 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
gpt-oss-120bMoEQ5_K_M120B58.57 GiB0.05 GiB59.51 GiB0.01 GiB98±37%
gpt-oss-safeguard-120bMoEQ5_K_M120B58.57 GiB0.05 GiB59.51 GiB0.01 GiB98±37%
llm-surgery-dark-arts-gpt-oss-60b-96a12MoEI1-Q6_K60.9B58.56 GiB0.03 GiB59.48 GiB0.04 GiB54±37%
gpt-oss-120b-uncensored-bf16MoEQ4_K_M117B58.53 GiB0.05 GiB59.46 GiB0.06 GiB98±37%
MiMo-V2-FlashMoEKV unresolvedI1-IQ1_S310B58.20 GiB0.13 GiB59.28 GiB0.24 GiB97±37%
Qwen3.5-REAP-212B-A17BMoEIQ2_XS212B58.30 GiB0.03 GiB59.28 GiB0.24 GiB94±37%
Step-3.7-FlashUD-IQ2_M201B57.58 GiB0.71 GiB59.22 GiB0.30 GiB18±26.5%
Hypernova-60B-2605MoEQ8_058.7B58.13 GiB0.04 GiB59.06 GiB0.46 GiB85±37%
Qwen3.5-88BMoEI1-Q5_K_M87.7B58.10 GiB0.03 GiB59.06 GiB0.46 GiB89±37%
command-a-plus-05-2026-bf16MoEIQ2_XXS219B57.98 GiB0.14 GiB59.02 GiB0.50 GiB77±37%
Qwen3.5-99BMoEI1-Q4_199.0B57.88 GiB0.03 GiB58.83 GiB0.69 GiB93±37%
Qwen3-235B-A22B-abliteratedMoEI1-IQ2_XXS235B57.55 GiB0.21 GiB58.69 GiB0.83 GiB77±37%
dots.llm1.instMoEQ2_K_L143B56.67 GiB1.09 GiB58.68 GiB0.84 GiB77±37%
gemma-4-31B-itBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma4-Gutenberg-31BBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Wanabi-Gemma4-31BBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma4-Gutenberg-31B-HereticBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
G4-MeroMero-31B-uncensored-hereticBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-31B-it-Claude-Opus-Distill-v2BF1632.7B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-31B-it-uncensored-hereticBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma4-31B-Fable-5-DistilledF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Glistening-Gem-31B-v1.0BF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma-4-31B-Fable-5-Agent-DistillBF1632.7B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
G4-MeroMero-31BBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
G4-Alice-v1.2-31BBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-hereticBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Equinox-31BBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma-4-Gembrain-31B-it-uncensored-hereticBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma-4-31B-Storymaxxed3F1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma-4-Harmonia-31B-uncensored-hereticBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-31B-MergemaxxedF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-31B-it-qat-q4_0-unquantized-hereticF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-31B-it-abliteratedF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma-4-Garnet-V2-31B-it-ultra-uncensored-hereticBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma-4-Queen-31B-it-uncensored-hereticBF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Gemma-4-31B-storymaxxed2F1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-31B-it-SDFT-Heretic-RPBF1630.7B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-31B-anthologyF1631.3B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
gemma-4-31BBF1632.7B57.20 GiB0.51 GiB58.68 GiB0.84 GiB18±26.5%
Qwen3.5-122B-A10BMoEUD-IQ4_XS125B57.67 GiB0.03 GiB58.63 GiB0.89 GiB100±37%
GLM-4.6VMoEQ4_0108B57.48 GiB0.20 GiB58.61 GiB0.91 GiB77±37%
Mixtral-8x22B-Instruct-v0.1MoEQ3_K_S141B57.28 GiB0.25 GiB58.49 GiB1.03 GiB34±37%
grok-2MoEIQ1_M270B57.16 GiB0.28 GiB58.49 GiB1.03 GiB34±37%
Mixtral-8x22B-v0.1MoEQ3_K_S141B57.28 GiB0.25 GiB58.49 GiB1.03 GiB34±37%
Mixtral-8x22B-v0.1MoEQ3_K_S141B57.27 GiB0.25 GiB58.48 GiB1.04 GiB34±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEUD-Q3_K_M124B57.47 GiB0.10 GiB58.47 GiB1.05 GiB90±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ4_K_S109B57.23 GiB0.21 GiB58.37 GiB1.15 GiB77±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_XXS236B57.28 GiB0.07 GiB58.29 GiB1.23 GiB94±37%
DeepSeek-V2.5MoEIQ2_XXS236B57.28 GiB0.07 GiB58.29 GiB1.23 GiB94±37%
DeepSeek-Coder-V2-InstructMoEIQ2_XXS236B57.28 GiB0.07 GiB58.29 GiB1.23 GiB94±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ3_M139B56.81 GiB0.27 GiB57.97 GiB1.55 GiB83±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ3_M139B56.81 GiB0.27 GiB57.97 GiB1.55 GiB83±37%
Devstral-2-123B-Instruct-2512Q3_K_M125B56.46 GiB0.39 GiB57.90 GiB1.62 GiB18±26.5%
Mistral-Medium-3.5-128BI1-Q3_K_M128B56.46 GiB0.39 GiB57.90 GiB1.62 GiB18±26.5%
XORTRON-NXTXPRTXXLI1-Q3_K_M128B56.46 GiB0.39 GiB57.90 GiB1.62 GiB18±26.5%
Qwen3-Coder-30B-A3B-InstructMoEBF1630.5B56.90 GiB0.11 GiB57.90 GiB1.62 GiB78±37%
Salience-1.5-FlashMoEBF1631.1B56.90 GiB0.11 GiB57.90 GiB1.62 GiB78±37%
Qwen3-VL-30B-A3B-InstructMoEBF1631.1B56.90 GiB0.11 GiB57.90 GiB1.62 GiB78±37%
Qwen3-VL-30B-A3B-ThinkingMoEBF1631.1B56.90 GiB0.11 GiB57.90 GiB1.62 GiB78±37%
MiroThinker-v1.0-30BMoEBF1630.5B56.90 GiB0.11 GiB57.90 GiB1.62 GiB78±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?
2063 of 2118 indexed open-weight models fit a Instinct MI210 at 4,096 context with q4_0 KV cache, the largest being gpt-oss-120b at Q5_K_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.