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. 2030 of 2118 indexed models fit at 32K context with q4_0 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 1743vision language 183video 16audio tts 21image 2embedding 26audio asr 39

What fits at 32K context

largest quantization that fits, per model · 2030 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Hypernova-60B-2605MoEI1-Q4_K_S58.7B35.89 GiB0.29 GiB37.17 GiB0.03 GiB115±37%
Mistral-Medium-3.5-128BIQ1_S128B32.92 GiB3.09 GiB37.17 GiB0.03 GiB25±22%
Gemma-2-9B-It-SPPO-Iter3F329.2B34.43 GiB1.68 GiB37.16 GiB0.04 GiB24±22%
gemma-2-9b-it-abliteratedF329.2B34.43 GiB1.68 GiB37.16 GiB0.04 GiB24±22%
Tiger-Gemma-9B-v1F329.2B34.43 GiB1.68 GiB37.16 GiB0.04 GiB24±22%
GLM-4.6VMoEUD-IQ1_S108B34.49 GiB1.62 GiB37.14 GiB0.06 GiB85±37%
Llama-4-Scout-17B-16E-Instruct-abliterated-v2MoEKV unresolvedI1-Q2_K_S109B34.42 GiB1.69 GiB37.14 GiB0.06 GiB84±37%
Qwen3.5-99BMoEI1-IQ3_XXS99.0B35.87 GiB0.21 GiB37.11 GiB0.09 GiB128±37%
Apertus-70B-Instruct-2509Q3_K_M70.6B33.10 GiB2.81 GiB37.09 GiB0.11 GiB25±22%
Rombo-LLM-V3.0-Qwen-72bI1-IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Qwen2.5-72B-Instruct-abliteratedI1-IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
HuatuoGPT-o1-72BIQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
MiroThinker-v1.0-72BI1-IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
EVA-Qwen2.5-72B-v0.2IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Qwen2.5-Math-72B-InstructIQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Qwen2.5-72B-InstructIQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Malaysian-Qwen2.5-72B-InstructI1-IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Qwen2.5-72BI1-IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
magnum-v4-72bI1-IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
KAT-Dev-72B-ExpIQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Homer-v1.0-Qwen2.5-72BIQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Qwen2.5-VL-72B-InstructIQ3_M73.4B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
Chronos-Platinum-72BIQ3_M72.7B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
UI-TARS-72B-DPOIQ3_M73.4B33.07 GiB2.81 GiB37.01 GiB0.19 GiB25±22%
c4ai-command-r-plus-08-2024IQ2_M104B33.56 GiB2.25 GiB36.99 GiB0.21 GiB25±22%
HarmonicHarlequin_v5-20BI1-Q4_K_S33.3B17.62 GiB18.28 GiB36.94 GiB0.26 GiB25±22%
Qwen3.5-88BMoEI1-IQ3_S87.7B35.64 GiB0.21 GiB36.88 GiB0.32 GiB123±37%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ3_K—35.65 GiB0.21 GiB36.85 GiB0.35 GiB155±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-Q3_K_M79.7B35.65 GiB0.21 GiB36.85 GiB0.35 GiB155±37%
Hunyuan-A13B-InstructMoEIQ3_M80.4B34.72 GiB1.13 GiB36.84 GiB0.36 GiB25±22%
Assistant_Pepe_70BQ3_K_M70.6B32.89 GiB2.81 GiB36.83 GiB0.37 GiB25±22%
CalmeRys-78B-Orpo-v0.1I1-IQ3_XS78.0B32.67 GiB3.02 GiB36.83 GiB0.37 GiB25±22%
Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4MoEQ8_034.7B35.60 GiB0.18 GiB36.78 GiB0.42 GiB140±37%
Ornith-Agents-A1-3.6-35B-A3B-dare_tiesMoEQ8_034.7B35.60 GiB0.18 GiB36.78 GiB0.42 GiB140±37%
Mistral-Small-4-119B-2603MoEIQ2_S119B35.52 GiB0.20 GiB36.75 GiB0.45 GiB139±37%
Llama-3_1-Nemotron-51B-InstructIQ2_XXS51.5B13.07 GiB22.50 GiB36.71 GiB0.49 GiB25±22%
Llama-3_3-Nemotron-Super-49B-v1_5UD-IQ2_XXS49.9B12.99 GiB22.50 GiB36.63 GiB0.57 GiB25±22%
Llama-3_3-Nemotron-Super-49B-v1UD-IQ2_XXS49.9B12.99 GiB22.50 GiB36.63 GiB0.57 GiB25±22%
Qwen3-Coder-Next-REAMMoEI1-Q4_160.3B35.30 GiB0.21 GiB36.51 GiB0.69 GiB146±37%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q5_K_S53.0B34.03 GiB1.48 GiB36.50 GiB0.70 GiB81±37%
GLM-4.5-Air-REAP-82B-A12BMoEQ2_K_L81.9B33.84 GiB1.62 GiB36.49 GiB0.71 GiB77±37%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEIQ4_XS35.1B35.30 GiB0.18 GiB36.48 GiB0.72 GiB141±37%
Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-PreservedQ8_027.4B34.80 GiB0.56 GiB36.42 GiB0.78 GiB25±22%
Salience-1.5-ProMoEQ8_036.0B35.22 GiB0.18 GiB36.40 GiB0.80 GiB141±37%
Qwable-v1MoEQ8_036.0B35.22 GiB0.18 GiB36.40 GiB0.80 GiB141±37%
T-SearchMoEQ8_036.0B35.22 GiB0.18 GiB36.40 GiB0.80 GiB141±37%
Qwen3.5-35B-A3BMoEQ8_036.0B35.22 GiB0.18 GiB36.40 GiB0.80 GiB141±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
Qwen3.6-35B-A3B-Fable-5-DistillMoEQ8_036.0B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
Qwable-v2MoEQ8_036.0B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
Qwen3.6-35B-A3B-YOYO-V2MoEQ8_036.0B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
Ornith-1.0-35B-FP8-BLOCK-MTPMoEQ8_035.5B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
fable-coder-35B-A3BMoEQ8_036.0B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
PINQWEN-3.6-35B-CLEAN-BF16MoEQ8_036.0B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
UniMath-35B-A3BMoEQ8_036.0B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
Ornith-1.0-35B-Heretic-MTPMoEQ8_0—35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
Fawen-1.0-35BMoEQ8_036.0B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±37%
Qwopus3.6-35B-A3B-v1MoEQ8_036.0B35.21 GiB0.18 GiB36.39 GiB0.81 GiB141±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?
2030 of 2118 indexed open-weight models fit a A100 40GB at 32,768 context with q4_0 KV cache, the largest being Hypernova-60B-2605 at I1-Q4_K_S. 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.