NVIDIA · datacenter

A100 40GB

A100 40GB has 40 GB of VRAM at 1555 GB/s — about 37.20 GiB usable after driver and compositor overhead. 2023 of 2118 indexed models fit at 16K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
40 GB
HBM2
Bandwidth
1555 GB/s
5120-bit bus
Tensor FP16
312 TF
dense
TDP
400 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1737vision language 182video 16audio tts 21image 2embedding 26audio asr 39

What fits at 16K context

largest quantization that fits, per model · 2023 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.5-Air-REAP-82B-A12BMoEQ2_K81.9B33.28 GiB2.88 GiB37.18 GiB0.02 GiB66±37%
llama-3.2-3b-instructF163.2B34.37 GiB1.75 GiB37.13 GiB0.07 GiB24±22%
CodeLlama-70b-Instruct-hfI1-Q3_K_M69.0B30.99 GiB5.00 GiB37.11 GiB0.09 GiB25±22%
CodeLlama-70b-Python-hfI1-Q3_K_M69.0B30.99 GiB5.00 GiB37.11 GiB0.09 GiB25±22%
Nous-Hermes-Llama2-70bI1-Q3_K_M69.0B30.99 GiB5.00 GiB37.11 GiB0.09 GiB25±22%
Midnight-Miqu-70B-v1.5I1-Q3_K_M69.0B30.99 GiB5.00 GiB37.11 GiB0.09 GiB25±22%
KafkaLM-70B-German-V0.1Q3_K_M69.0B30.99 GiB5.00 GiB37.11 GiB0.09 GiB25±22%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-IQ2_M109B33.04 GiB3.00 GiB37.06 GiB0.14 GiB71±37%
Qwen3.5-88BMoEI1-IQ3_S87.7B35.64 GiB0.38 GiB37.04 GiB0.16 GiB119±37%
llama2_70b_chat_uncensoredQ3_K_M69.0B30.91 GiB5.00 GiB37.03 GiB0.17 GiB25±22%
Xwin-LM-70b-V0.1Q3_K_M69.0B30.91 GiB5.00 GiB37.03 GiB0.17 GiB25±22%
Llama-2-70b-chat-hfQ3_K_M69.0B30.91 GiB5.00 GiB37.03 GiB0.17 GiB25±22%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ3_K—35.65 GiB0.38 GiB37.01 GiB0.19 GiB147±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-Q3_K_M79.7B35.65 GiB0.38 GiB37.01 GiB0.19 GiB147±37%
OLMo-2-0325-32BQ8_032.2B31.90 GiB4.00 GiB37.00 GiB0.20 GiB25±22%
GLM-4.5VMoEI1-IQ1_M108B33.08 GiB2.88 GiB36.98 GiB0.22 GiB72±37%
Wizard-Vicuna-30B-UncensoredI1-IQ3_XXS32.5B11.48 GiB24.38 GiB36.92 GiB0.28 GiB25±22%
archangel_sft-kto_llama30bI1-IQ3_XXS32.5B11.48 GiB24.38 GiB36.92 GiB0.28 GiB25±22%
Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4MoEQ8_034.7B35.60 GiB0.31 GiB36.91 GiB0.29 GiB135±37%
Ornith-Agents-A1-3.6-35B-A3B-dare_tiesMoEQ8_034.7B35.60 GiB0.31 GiB36.91 GiB0.29 GiB135±37%
Mistral-Small-4-119B-2603MoEIQ2_S119B35.52 GiB0.35 GiB36.90 GiB0.30 GiB134±37%
Delphi-25B-SimpleRL-MathI1-Q6_K25.0B19.08 GiB16.73 GiB36.89 GiB0.31 GiB25±22%
Mixtral_34Bx2_MoE_60BMoEQ4_K_M60.8B32.03 GiB3.75 GiB36.86 GiB0.34 GiB14±37%
Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-PreservedQ8_027.4B34.80 GiB1.00 GiB36.86 GiB0.34 GiB25±22%
Behemoth-X-123B-v2IQ2_XXS123B30.20 GiB5.50 GiB36.86 GiB0.34 GiB25±22%
Mistral-Large-Instruct-2411IQ2_XXS123B30.20 GiB5.50 GiB36.86 GiB0.34 GiB25±22%
Assistant_Pepe_70BIQ3_M70.6B30.72 GiB5.00 GiB36.85 GiB0.35 GiB25±22%
Mistral-Small-Instruct-2409Q2_K22.2B32.27 GiB3.50 GiB36.83 GiB0.37 GiB25±22%
Rombo-LLM-V3.0-Qwen-72bI1-IQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
Qwen2.5-72B-Instruct-abliteratedI1-IQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
MiroThinker-v1.0-72BI1-IQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
Malaysian-Qwen2.5-72B-InstructI1-IQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
Qwen2.5-72BI1-IQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
magnum-v4-72bI1-IQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
KAT-Dev-72B-ExpIQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ3_XS72.7B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
Qwen2.5-VL-72B-InstructIQ3_XS73.4B30.59 GiB5.00 GiB36.72 GiB0.48 GiB25±22%
Qwen3-Coder-Next-REAMMoEI1-Q4_160.3B35.30 GiB0.38 GiB36.67 GiB0.53 GiB140±37%
command-r-35b-writer-v2I1-IQ3_M35.0B15.55 GiB20.00 GiB36.66 GiB0.54 GiB25±22%
Laguna-S-2.1MoEUD-IQ2_M118B34.71 GiB0.89 GiB36.62 GiB0.58 GiB112±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEIQ4_XS35.1B35.30 GiB0.31 GiB36.62 GiB0.58 GiB136±37%
Qwen3-Coder-NextMoEIQ3_M79.7B34.13 GiB1.50 GiB36.62 GiB0.58 GiB114±37%
Qwen3-Next-80B-A3B-ThinkingMoEIQ3_M81.3B34.13 GiB1.50 GiB36.62 GiB0.58 GiB114±37%
Qwen3-Next-80B-A3B-InstructMoEIQ3_M81.3B34.13 GiB1.50 GiB36.62 GiB0.58 GiB114±37%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-IQ1_M109B32.59 GiB3.00 GiB36.61 GiB0.59 GiB71±37%
Gemma-4-31B-Isometry-RPQ8_032.7B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
Prosopon-31BQ8_032.7B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
Gemma-4-Novelist-Eclipse-31BQ8_032.7B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
Giftige-Blume-31B-v1-StyleSwapQ8_032.7B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
G4-MeroMero-31B-StyleSwapQ8_032.7B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
Gemma-4-31B-StyleTune-heretic-araQ8_032.7B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
Pantheon-Reasoning-31B-1.1Q8_032.7B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
Gemma-4-31B-StyleTuneQ8_032.7B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
Barcenas-StyleTune-31B-FableQ8_032.1B31.79 GiB3.67 GiB36.54 GiB0.66 GiB25±22%
Salience-1.5-ProMoEQ8_036.0B35.22 GiB0.31 GiB36.53 GiB0.67 GiB136±37%
Qwable-v1MoEQ8_036.0B35.22 GiB0.31 GiB36.53 GiB0.67 GiB136±37%
T-SearchMoEQ8_036.0B35.22 GiB0.31 GiB36.53 GiB0.67 GiB136±37%
Qwen3.5-35B-A3BMoEQ8_036.0B35.22 GiB0.31 GiB36.53 GiB0.67 GiB136±37%
Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-UncensoredMoEQ8_033.6B33.27 GiB2.25 GiB36.53 GiB0.67 GiB52±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 A100 40GB run?
2023 of 2118 indexed open-weight models fit a A100 40GB at 16,384 context with f16 KV cache, the largest being GLM-4.5-Air-REAP-82B-A12B at Q2_K. That covers text, vision-language, image, video and speech models.
How much usable memory does a A100 40GB actually have?
Its nameplate is 40 GB, but about 37.20 GiB is available to a model once driver and compositor overhead is accounted for.
Is a A100 40GB fast for local AI?
Its memory bandwidth is 1555 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.