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 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 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◐
Hunyuan-A13B-InstructMoEQ3_K_M80.4B35.91 GiB0.28 GiB37.19 GiB0.01 GiB24±22%
Qwen3.5-88BMoEI1-IQ3_M87.7B36.10 GiB0.05 GiB37.18 GiB0.02 GiB127±37%
Hermes-4-70BIQ4_XS70.6B35.33 GiB0.70 GiB37.16 GiB0.04 GiB25±22%
Llama-3.3-70B-InstructIQ4_XS70.6B35.33 GiB0.70 GiB37.16 GiB0.04 GiB25±22%
DeepSeek-R1-Distill-Llama-70BIQ4_XS70.6B35.33 GiB0.70 GiB37.16 GiB0.04 GiB25±22%
Maenad-70BI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
calme-2.4-llama3-70bIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
calme-2.2-llama3-70bIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Rombos-LLM-70b-Llama-3.3I1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
L3.3-Electra-R1-70bI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
L3.3-70B-Magnum-v4-SEIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Latxa-Llama-3.1-70B-Instruct-v2I1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Llama-3.3_70_b_uncensored_continuedI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Llama-3.3-70B-Instruct-abliteratedI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
grok-oss-Revenant-70BI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Llama-3.1-Nemotron-70B-Instruct-HFI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
L3.3-70B-Euryale-v2.3I1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Hermes-3-Llama-3.1-70BIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Hermes-4-70B-hereticI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Llama-3.1-70BIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Anubis-70B-v1.2IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Golem-70B-v1bI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-hereticI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
llama-3-firefunction-v2IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Legion-V2.1-LLaMa-70BI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Assistant_Pepe_70BI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Tess-R1-Limerick-Llama-3.1-70BIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
SEMIKONG-70BIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
functionary-medium-v3.2KV unresolvedIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Llama-3.1-WhiteRabbitNeo-2-70BIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
New-Dawn-Llama-3-70B-32K-v1.0I1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-IQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Athene-70BIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
L3.3-70B-Magnum-DiamondIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
Meta-Llama-3-70B-InstructIQ4_XS70.6B35.30 GiB0.70 GiB37.13 GiB0.07 GiB25±22%
xLAM-8x7b-rMoEQ6_K_L46.7B35.80 GiB0.28 GiB37.12 GiB0.08 GiB46±37%
Qwen3.5-122B-A10BMoEUD-IQ1_M125B36.02 GiB0.05 GiB37.10 GiB0.10 GiB143±37%
GLM-4.5-AirMoEUD-TQ1_0110B35.63 GiB0.40 GiB37.07 GiB0.13 GiB104±37%
dolphin-2.6-mixtral-8x7bMoEI1-Q6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
Nous-Hermes-2-Mixtral-8x7B-DPOMoEQ6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
Mixtral-8x7B-Instruct-v0.1MoEQ6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
Open_Gpt4_8x7B_v0.1MoEQ6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
dolphin-2.5-mixtral-8x7bMoEQ6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
dolphin-2.7-mixtral-8x7bMoEQ6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
Mixtral-8x7B-v0.1MoEQ6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
Mixtral-8x7B-MoE-RP-StoryMoEQ6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
Open_Gpt4_8x7B_v0.2MoEQ6_K46.7B35.74 GiB0.28 GiB37.06 GiB0.14 GiB46±37%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-Q5_K_S57.3B35.72 GiB0.29 GiB37.05 GiB0.15 GiB106±37%
GLM-4.5VMoEI1-IQ2_XXS108B35.61 GiB0.40 GiB37.04 GiB0.16 GiB104±37%
Qwen3-Coder-NextMoEUD-IQ4_XS79.7B35.79 GiB0.21 GiB36.99 GiB0.21 GiB154±37%
Hypernova-60B-2605MoEI1-Q4_K_S58.7B35.89 GiB0.08 GiB36.96 GiB0.24 GiB121±37%
Qwen3.5-99BMoEI1-IQ3_XXS99.0B35.87 GiB0.05 GiB36.95 GiB0.25 GiB134±37%
Llama-3_1-Nemotron-51B-InstructQ4_151.5B30.18 GiB5.63 GiB36.95 GiB0.25 GiB25±22%
Rombo-LLM-V3.0-Qwen-72bI1-Q3_K_M72.7B35.11 GiB0.70 GiB36.94 GiB0.26 GiB25±22%
Qwen2.5-72B-Instruct-abliteratedI1-Q3_K_M72.7B35.11 GiB0.70 GiB36.94 GiB0.26 GiB25±22%
Qwen2.5-72B-Instruct-abliterated-v2I1-Q3_K_M72.7B35.11 GiB0.70 GiB36.94 GiB0.26 GiB25±22%
HuatuoGPT-o1-72BQ3_K_M72.7B35.11 GiB0.70 GiB36.94 GiB0.26 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?
2032 of 2118 indexed open-weight models fit a A100 40GB at 8,192 context with q4_0 KV cache, the largest being Hunyuan-A13B-Instruct at Q3_K_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.