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. 2032 of 2118 indexed models fit at 8K 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 1745vision language 183video 16audio tts 21image 2embedding 26audio asr 39

What fits at 8K context

largest quantization that fits, per model · 2032 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Qwen3-Coder-NextMoEUD-IQ4_XS79.7B35.79 GiB0.40 GiB37.18 GiB0.02 GiB146±37%
gemma-4-31B-it-abliteratedQ4_K_M31.3B34.81 GiB1.29 GiB37.17 GiB0.03 GiB25±22%
Qwen3.5-122B-A10BMoEUD-IQ1_M125B36.02 GiB0.10 GiB37.14 GiB0.06 GiB141±37%
CodeLlama-70b-Python-hfIQ4_XS69.0B34.64 GiB1.33 GiB37.09 GiB0.11 GiB25±22%
Midnight-Miqu-70B-v1.5IQ4_XS69.0B34.64 GiB1.33 GiB37.09 GiB0.11 GiB25±22%
Qwen3-Next-80B-A3B-ThinkingMoEQ3_K_M81.3B35.67 GiB0.40 GiB37.06 GiB0.14 GiB146±37%
Qwen3-Next-80B-A3B-InstructMoEQ3_K_M81.3B35.67 GiB0.40 GiB37.06 GiB0.14 GiB146±37%
Meta-Llama-3-70B-InstructQ3_K_L70.6B34.60 GiB1.33 GiB37.05 GiB0.15 GiB25±22%
GLM-4.5-Air-DerestrictedMoEIQ1_M110B35.26 GiB0.76 GiB37.05 GiB0.15 GiB97±37%
GLM-4.5-AirMoEIQ1_M110B35.26 GiB0.76 GiB37.05 GiB0.15 GiB97±37%
Maenad-70BI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
calme-2.4-llama3-70bQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
calme-2.2-llama3-70bQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Rombos-LLM-70b-Llama-3.3I1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
L3.3-Electra-R1-70bI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
L3.3-70B-Magnum-v4-SEQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Llama-3.3_70_b_uncensored_continuedI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Llama-3.3-70B-Instruct-abliteratedI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Strawberrylemonade-L3-70B-v1.2Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
grok-oss-Revenant-70BI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
L3.3-70B-Euryale-v2.3I1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Hermes-4-70B-hereticI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Llama-3.3-70B-InstructQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Llama-3.1-70BQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Hermes-3-Llama-3.1-70BQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Anubis-70B-v1.2Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Hermes-4-70BQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Golem-70B-v1bI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
llama-3-firefunction-v2Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
DeepSeek-R1-Distill-Llama-70BQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Legion-V2.1-LLaMa-70BI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Assistant_Pepe_70BI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Tess-R1-Limerick-Llama-3.1-70BQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
SEMIKONG-70BQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
functionary-medium-v3.2KV unresolvedQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Llama-3.1-WhiteRabbitNeo-2-70BQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
New-Dawn-Llama-3-70B-32K-v1.0I1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-Q3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Athene-70BQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
L3.3-70B-Magnum-DiamondQ3_K_L70.6B34.59 GiB1.33 GiB37.04 GiB0.16 GiB25±22%
Hypernova-60B-2605MoEI1-Q4_K_S58.7B35.89 GiB0.15 GiB37.03 GiB0.17 GiB119±37%
Qwen3.5-99BMoEI1-IQ3_XXS99.0B35.87 GiB0.10 GiB37.00 GiB0.20 GiB132±37%
GLM-4.6VMoEIQ2_XXS108B35.11 GiB0.76 GiB36.90 GiB0.30 GiB98±37%
CalmeRys-78B-Orpo-v0.1I1-IQ3_S78.0B34.33 GiB1.43 GiB36.89 GiB0.31 GiB25±22%
Qwen2.5-Coder-32B-InstructQ4_032.8B34.72 GiB1.06 GiB36.88 GiB0.32 GiB25±22%
Llama-3_3-Nemotron-Super-49B-v1_5IQ4_XS49.9B25.06 GiB10.63 GiB36.82 GiB0.38 GiB25±22%
Llama-3_3-Nemotron-Super-49B-v1IQ4_XS49.9B25.06 GiB10.63 GiB36.82 GiB0.38 GiB25±22%
calme-2.3-rys-78bQ3_K_S78.0B34.24 GiB1.43 GiB36.80 GiB0.40 GiB25±22%
Valkyrie-49B-v2.1I1-IQ4_XS49.9B25.03 GiB10.63 GiB36.79 GiB0.41 GiB25±22%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q5_K_M53.0B35.08 GiB0.70 GiB36.77 GiB0.43 GiB90±37%
Qwen3.5-88BMoEI1-IQ3_S87.7B35.64 GiB0.10 GiB36.77 GiB0.43 GiB127±37%
CodeLlama-70b-Instruct-hfI1-IQ4_XS69.0B34.30 GiB1.33 GiB36.75 GiB0.45 GiB25±22%
Nous-Hermes-Llama2-70bI1-IQ4_XS69.0B34.30 GiB1.33 GiB36.75 GiB0.45 GiB25±22%
Qwen3-Coder-Next-Opus-4.6-Reasoning-DistilledMoEQ3_K—35.65 GiB0.10 GiB36.74 GiB0.46 GiB160±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?
2032 of 2118 indexed open-weight models fit a A100 40GB at 8,192 context with q8_0 KV cache, the largest being Qwen3-Coder-Next at UD-IQ4_XS. 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.