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. 2056 of 2118 indexed models fit at 64K 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 1767vision language 185image 2audio asr 39audio tts 21video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 2056 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
HarmonicHarlequin_v5-20BI1-Q5_K_M33.3B21.98 GiB36.56 GiB59.49 GiB0.03 GiB18±26.5%
Qwen3-Coder-30B-A3B-InstructMoEBF1630.5B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Salience-1.5-FlashMoEBF1631.1B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Qwen3-VL-30B-A3B-InstructMoEBF1631.1B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Qwen3-VL-30B-A3B-ThinkingMoEBF1631.1B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
MiroThinker-v1.0-30BMoEBF1630.5B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Qwen3-30B-A3BMoEBF1630.5B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Qwen3-30B-A3B-Thinking-2507-Claude-4.5-Sonnet-High-Reasoning-DistillMoEBF1630.5B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Qwen3-30B-A3B-Instruct-2507MoEBF1630.5B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Qwen3-30B-A3B-Thinking-2507MoEBF1630.5B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Pantheon-Proto-RP-1.8-30B-A3BMoEBF1630.5B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Tongyi-DeepResearch-30B-A3BMoEBF1630.5B56.90 GiB1.69 GiB59.48 GiB0.04 GiB65±37%
Qwen3.5-88BMoEI1-Q5_K_M87.7B58.10 GiB0.42 GiB59.45 GiB0.07 GiB84±37%
DeepSeek-Coder-V2-Instruct-0724MoEIQ2_XXS236B57.28 GiB1.19 GiB59.40 GiB0.12 GiB81±37%
DeepSeek-V2.5MoEIQ2_XXS236B57.28 GiB1.19 GiB59.40 GiB0.12 GiB81±37%
DeepSeek-Coder-V2-InstructMoEIQ2_XXS236B57.28 GiB1.19 GiB59.40 GiB0.12 GiB81±37%
NVIDIA-Nemotron-3-Super-120B-A12B-BF16MoEQ3_K_S124B56.93 GiB1.55 GiB59.37 GiB0.15 GiB74±37%
step-3.5-flashIQ2_XS199B51.40 GiB7.04 GiB59.37 GiB0.15 GiB18±26.5%
Qwen2.5-14B-Instruct-1MF3214.8B55.03 GiB3.38 GiB59.35 GiB0.17 GiB18±26.5%
DeepSeek-R1-Distill-Qwen-14BF3214.8B55.03 GiB3.38 GiB59.35 GiB0.17 GiB18±26.5%
granite-4.1-30bBF1628.9B53.77 GiB4.50 GiB59.28 GiB0.24 GiB18±26.5%
Qwen3.5-99BMoEI1-Q4_199.0B57.88 GiB0.42 GiB59.23 GiB0.29 GiB88±37%
Step-3.7-FlashIQ2_XXS201B51.22 GiB7.04 GiB59.19 GiB0.33 GiB18±26.5%
Mixtral-8x22B-Instruct-v0.1MoEIQ3_XS141B54.23 GiB3.94 GiB59.13 GiB0.39 GiB29±37%
Mixtral-8x22B-v0.1MoEIQ3_XS141B54.23 GiB3.94 GiB59.13 GiB0.39 GiB29±37%
Mixtral-8x22B-v0.1MoEIQ3_XS141B54.23 GiB3.94 GiB59.13 GiB0.39 GiB29±37%
Llama-3_1-Nemotron-51B-InstructIQ2_XXS51.5B13.07 GiB45.00 GiB59.11 GiB0.41 GiB18±26.5%
phi-4F3214.7B54.61 GiB3.52 GiB59.09 GiB0.43 GiB18±26.5%
Llama-3_3-Nemotron-Super-49B-v1_5UD-IQ2_XXS49.9B12.99 GiB45.00 GiB59.03 GiB0.49 GiB18±26.5%
Llama-3_3-Nemotron-Super-49B-v1UD-IQ2_XXS49.9B12.99 GiB45.00 GiB59.03 GiB0.49 GiB18±26.5%
Qwen3.5-122B-A10BMoEUD-IQ4_XS125B57.67 GiB0.42 GiB59.02 GiB0.50 GiB94±37%
GLM-4.7-REAP-218B-A32BMoEIQ2_XXS218B51.57 GiB6.47 GiB58.98 GiB0.54 GiB42±37%
GLM-4.6VMoEIQ4_XS108B54.80 GiB3.23 GiB58.97 GiB0.55 GiB57±37%
Qwen3-14B-GPT-5.2-High-Reasoning-DistillBF1614.8B55.03 GiB2.81 GiB58.80 GiB0.72 GiB18±26.5%
Valkyrie-49B-v2.1I1-IQ2_XXS49.9B12.72 GiB45.00 GiB58.76 GiB0.76 GiB18±26.5%
Behemoth-X-123B-v2IQ3_M123B51.48 GiB6.19 GiB58.72 GiB0.80 GiB18±26.5%
Mistral-Large-Instruct-2411IQ3_M123B51.48 GiB6.19 GiB58.72 GiB0.80 GiB18±26.5%
Hy-MT2-30B-A3BMoEBF1630.1B56.03 GiB1.69 GiB58.61 GiB0.91 GiB66±37%
openPangu-2.0-FlashMoEKV unresolvedQ4_K_M100B56.71 GiB0.91 GiB58.53 GiB0.99 GiB88±37%
Mistral-Medium-3.5-128BIQ3_XS128B51.16 GiB6.19 GiB58.40 GiB1.12 GiB18±26.5%
Huihui-GLM-4.7-Flash-abliterated-57BMoEQ8_057.3B55.03 GiB2.35 GiB58.33 GiB1.19 GiB62±37%
GLM-4.5VMoEI1-IQ4_XS108B54.13 GiB3.23 GiB58.29 GiB1.23 GiB58±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ3_XS139B53.00 GiB4.36 GiB58.24 GiB1.28 GiB55±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ3_XS139B53.00 GiB4.36 GiB58.24 GiB1.28 GiB55±37%
North-Mini-Code-1.0MoEBF1630.5B56.81 GiB0.55 GiB58.24 GiB1.28 GiB74±37%
command-r-35b-writer-v2Q8_035.0B34.63 GiB22.50 GiB58.13 GiB1.39 GiB18±26.5%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedQ3_K_L109B53.83 GiB3.38 GiB58.13 GiB1.39 GiB57±37%
XORTRON-NXTXPRTXXLI1-IQ3_S128B50.77 GiB6.19 GiB58.01 GiB1.51 GiB18±26.5%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPQ8_027.8B55.90 GiB1.13 GiB57.99 GiB1.53 GiB18±26.5%
c4ai-command-r-plus-08-2024IQ4_XS104B52.34 GiB4.50 GiB57.92 GiB1.60 GiB18±26.5%
Devstral-2-123B-Instruct-2512Q3_K_S125B50.63 GiB6.19 GiB57.88 GiB1.64 GiB18±26.5%
Qwen3-Omni-30B-A3B-InstructBF1635.3B56.90 GiB0.00 GiB57.85 GiB1.67 GiB18±26.5%
Qwen3-Omni-30B-A3B-ThinkingBF1631.7B56.90 GiB0.00 GiB57.85 GiB1.67 GiB18±26.5%
InternVL3_5-30B-A3BBF1630.8B56.90 GiB0.00 GiB57.84 GiB1.68 GiB18±26.5%
Qwen3-Coder-NextMoEUD-Q5_K_M79.7B55.17 GiB1.69 GiB57.75 GiB1.77 GiB88±37%
MiniMax-M2MoEUD-TQ1_0229B52.50 GiB4.36 GiB57.74 GiB1.78 GiB61±37%
Qwen3.8-27BBF1627.8B55.65 GiB1.13 GiB57.74 GiB1.78 GiB18±26.5%
GLM-4.7-FlashMoEBF1631.2B55.79 GiB0.93 GiB57.63 GiB1.89 GiB72±37%
GLM-4.7-Flash-Claude-Opus-4.5-High-Reasoning-DistillMoEBF1631.2B55.79 GiB0.93 GiB57.63 GiB1.89 GiB72±37%
GLM-4.7-Flash-hereticMoEBF1629.9B55.79 GiB0.93 GiB57.63 GiB1.89 GiB72±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?
2056 of 2118 indexed open-weight models fit a Instinct MI210 at 65,536 context with q4_0 KV cache, the largest being HarmonicHarlequin_v5-20B at I1-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.