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 32K context with q8_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 1737vision language 182video 16audio tts 21image 2embedding 26audio asr 39

What fits at 32K context

largest quantization that fits, per model · 2023 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
GLM-4.5VMoEI1-IQ1_M108B33.08 GiB3.05 GiB37.16 GiB0.04 GiB70±37%
Assistant_Pepe_70BIQ3_M70.6B30.72 GiB5.31 GiB37.16 GiB0.04 GiB25±22%
command-r-35b-writer-v2I1-IQ3_S35.0B14.77 GiB21.25 GiB37.13 GiB0.07 GiB25±22%
Mixtral_34Bx2_MoE_60BMoEQ4_K_M60.8B32.03 GiB3.98 GiB37.10 GiB0.10 GiB14±37%
Qwen3.5-88BMoEI1-IQ3_S87.7B35.64 GiB0.40 GiB37.07 GiB0.13 GiB118±37%
Mistral-Small-Instruct-2409Q2_K22.2B32.27 GiB3.72 GiB37.05 GiB0.15 GiB25±22%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ3_K—35.65 GiB0.40 GiB37.04 GiB0.16 GiB146±37%
Huihui-Qwen3-Coder-Next-abliteratedMoEI1-Q3_K_M79.7B35.65 GiB0.40 GiB37.03 GiB0.17 GiB146±37%
Rombo-LLM-V3.0-Qwen-72bI1-IQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
Qwen2.5-72B-Instruct-abliteratedI1-IQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
Qwen2.5-72B-Instruct-abliterated-v2I1-IQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
MiroThinker-v1.0-72BI1-IQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
Malaysian-Qwen2.5-72B-InstructI1-IQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
Qwen2.5-72BI1-IQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
magnum-v4-72bI1-IQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
KAT-Dev-72B-ExpIQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
Tower-Plus-72B-ultra-uncensored-hereticI1-IQ3_XS72.7B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
Qwen2.5-VL-72B-InstructIQ3_XS73.4B30.59 GiB5.31 GiB37.03 GiB0.17 GiB25±22%
Ornith-Agents-A1-3.7-35B-A3B-dare_ties_v4MoEQ8_034.7B35.60 GiB0.33 GiB36.93 GiB0.27 GiB134±37%
Ornith-Agents-A1-3.6-35B-A3B-dare_tiesMoEQ8_034.7B35.60 GiB0.33 GiB36.93 GiB0.27 GiB134±37%
Mistral-Small-4-119B-2603MoEIQ2_S119B35.52 GiB0.37 GiB36.92 GiB0.28 GiB133±37%
Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-PreservedQ8_027.4B34.80 GiB1.06 GiB36.92 GiB0.28 GiB25±22%
Noromaid-20b-v0.1.1I1-Q6_K20.0B15.28 GiB20.59 GiB36.91 GiB0.29 GiB25±22%
Nethena-20BQ6_K20.0B15.28 GiB20.59 GiB36.91 GiB0.29 GiB25±22%
IQuest-Coder-V1-40B-InstructI1-Q6_K39.8B30.41 GiB5.31 GiB36.82 GiB0.38 GiB25±22%
Llama-4-Scout-17B-16E-InstructMoEKV unresolvedUD-IQ1_M109B32.59 GiB3.19 GiB36.80 GiB0.40 GiB69±37%
Qwen3-Coder-NextMoEIQ3_M79.7B34.13 GiB1.59 GiB36.71 GiB0.49 GiB112±37%
Qwen3-Next-80B-A3B-ThinkingMoEIQ3_M81.3B34.13 GiB1.59 GiB36.71 GiB0.49 GiB112±37%
Qwen3-Next-80B-A3B-InstructMoEIQ3_M81.3B34.13 GiB1.59 GiB36.71 GiB0.49 GiB112±37%
Qwen3-Coder-Next-REAMMoEI1-Q4_160.3B35.30 GiB0.40 GiB36.69 GiB0.51 GiB139±37%
GLM-4.5-Air-REAP-82B-A12BMoEIQ2_M81.9B32.59 GiB3.05 GiB36.68 GiB0.52 GiB65±37%
Qwen3-48B-A4B-Savant-Commander-Distill-12X-Closed-Open-Heretic-UncensoredMoEQ8_033.6B33.27 GiB2.39 GiB36.67 GiB0.53 GiB52±37%
Wizard-Vicuna-30B-UncensoredI1-IQ2_S32.5B9.67 GiB25.90 GiB36.64 GiB0.56 GiB25±22%
archangel_sft-kto_llama30bI1-IQ2_S32.5B9.67 GiB25.90 GiB36.64 GiB0.56 GiB25±22%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEIQ4_XS35.1B35.30 GiB0.33 GiB36.64 GiB0.56 GiB135±37%
Laguna-S-2.1MoEUD-IQ2_M118B34.71 GiB0.87 GiB36.60 GiB0.60 GiB113±37%
Salience-1.5-ProMoEQ8_036.0B35.22 GiB0.33 GiB36.55 GiB0.65 GiB135±37%
Qwable-v1MoEQ8_036.0B35.22 GiB0.33 GiB36.55 GiB0.65 GiB135±37%
T-SearchMoEQ8_036.0B35.22 GiB0.33 GiB36.55 GiB0.65 GiB135±37%
Qwen3.5-35B-A3BMoEQ8_036.0B35.22 GiB0.33 GiB36.55 GiB0.65 GiB135±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Qwen3.6-35B-A3B-Fable-5-DistillMoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Qwable-v2MoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Qwen3.6-35B-A3B-YOYO-V2MoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Ornith-1.0-35B-FP8-BLOCK-MTPMoEQ8_035.5B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
fable-coder-35B-A3BMoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
PINQWEN-3.6-35B-CLEAN-BF16MoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
UniMath-35B-A3BMoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Ornith-1.0-35B-Heretic-MTPMoEQ8_0—35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Fawen-1.0-35BMoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Qwopus3.6-35B-A3B-v1MoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
CyberStrike-OffSec-35BMoEQ8_035.1B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Qwen3.6-35B-A3BMoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-DistilledMoEQ8_036.0B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Qwen3.5-35B-A3B-uncensored-heretic-v2-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.33 GiB36.54 GiB0.66 GiB135±37%
Melody1437-27BQ4_K_M27.8B34.41 GiB1.06 GiB36.54 GiB0.66 GiB25±22%
CalmeRys-78B-Orpo-v0.1I1-Q2_K78.0B29.66 GiB5.71 GiB36.50 GiB0.70 GiB25±22%
calme-2.3-rys-78bQ2_K78.0B29.66 GiB5.71 GiB36.50 GiB0.70 GiB25±22%
c4ai-command-r-plus-08-2024IQ2_S104B31.04 GiB4.25 GiB36.47 GiB0.73 GiB25±22%
Apertus-70B-Instruct-2509IQ3_M70.6B29.84 GiB5.31 GiB36.33 GiB0.87 GiB25±22%
●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 32,768 context with q8_0 KV cache, the largest being GLM-4.5V at I1-IQ1_M. 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.