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. 2001 of 2118 indexed models fit at 128K 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 1717vision language 180video 16audio tts 21audio asr 39image 2embedding 26

What fits at 128K context

largest quantization that fits, per model · 2001 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
NSFW_13B_sftQ4_K_M13.3B7.97 GiB28.13 GiB37.14 GiB0.06 GiB25±22%
Qwen3-Coder-Next-REAMMoEI1-Q4_160.3B35.30 GiB0.84 GiB37.14 GiB0.06 GiB124±37%
CodeLlama-70b-Instruct-hfI1-IQ3_XXS69.0B24.76 GiB11.25 GiB37.13 GiB0.07 GiB25±22%
CodeLlama-70b-Python-hfI1-IQ3_XXS69.0B24.76 GiB11.25 GiB37.13 GiB0.07 GiB25±22%
Nous-Hermes-Llama2-70bI1-IQ3_XXS69.0B24.76 GiB11.25 GiB37.13 GiB0.07 GiB25±22%
Midnight-Miqu-70B-v1.5I1-IQ3_XXS69.0B24.76 GiB11.25 GiB37.13 GiB0.07 GiB25±22%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-Q4_K_M57.3B31.37 GiB4.71 GiB37.12 GiB0.08 GiB59±37%
Qwen2.5-Coder-14B-InstructQ8_014.8B29.25 GiB6.75 GiB37.04 GiB0.16 GiB25±22%
Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16MoEIQ4_XS35.1B35.30 GiB0.70 GiB37.01 GiB0.19 GiB123±37%
Qwen3.6-34B-80L-Fable-5-HereticQ8_033.4B33.08 GiB2.81 GiB36.95 GiB0.25 GiB25±22%
OLMo-2-1124-13B-InstructQ4_K_M13.7B7.78 GiB28.13 GiB36.95 GiB0.25 GiB25±22%
Maenad-70BI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
calme-2.4-llama3-70bQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
calme-2.2-llama3-70bQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Rombos-LLM-70b-Llama-3.3I1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
L3.3-Electra-R1-70bI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
L3.3-70B-Magnum-v4-SEQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Latxa-Llama-3.1-70B-Instruct-v2I1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Llama-3.3_70_b_uncensored_continuedI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Llama-3.3-70B-Instruct-abliteratedI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Strawberrylemonade-L3-70B-v1.2Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
grok-oss-Revenant-70BI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Llama-3.1-Nemotron-70B-Instruct-HFI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
L3.3-70B-Euryale-v2.3I1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Hermes-4-70B-hereticI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Hermes-4-70BQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Llama-3.3-70B-InstructQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Llama-3.1-70BQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Anubis-70B-v1.2Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Golem-70B-v1bI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
DeepSeek-R1-Distill-Llama-70B-hereticI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
DeepSeek-R1-Distill-Llama-70BQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
llama-3-firefunction-v2Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Legion-V2.1-LLaMa-70BI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Assistant_Pepe_70BI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Tess-R1-Limerick-Llama-3.1-70BQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
SEMIKONG-70BQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
functionary-medium-v3.2KV unresolvedQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Llama-3.1-WhiteRabbitNeo-2-70BQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Infinity-Instruct-7M-Gen-Llama3_1-70BI1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
New-Dawn-Llama-3-70B-32K-v1.0I1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Meta-Llama-3-70B-Instruct-abliterated-v3.5I1-Q2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Athene-70BQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Hermes-3-Llama-3.1-70BQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
L3.3-70B-Magnum-DiamondQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Meta-Llama-3-70B-InstructQ2_K70.6B24.56 GiB11.25 GiB36.94 GiB0.26 GiB25±22%
Salience-1.5-ProMoEQ8_036.0B35.22 GiB0.70 GiB36.92 GiB0.28 GiB123±37%
Qwable-v1MoEQ8_036.0B35.22 GiB0.70 GiB36.92 GiB0.28 GiB123±37%
T-SearchMoEQ8_036.0B35.22 GiB0.70 GiB36.92 GiB0.28 GiB123±37%
Qwen3.5-35B-A3BMoEQ8_036.0B35.22 GiB0.70 GiB36.92 GiB0.28 GiB123±37%
Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-PreservedMoEQ8_035.1B35.21 GiB0.70 GiB36.91 GiB0.29 GiB123±37%
Qwen3.6-35B-A3B-Fable-5-DistillMoEQ8_036.0B35.21 GiB0.70 GiB36.91 GiB0.29 GiB123±37%
Qwable-v2MoEQ8_036.0B35.21 GiB0.70 GiB36.91 GiB0.29 GiB123±37%
Qwen3.6-35B-A3B-YOYO-V2MoEQ8_036.0B35.21 GiB0.70 GiB36.91 GiB0.29 GiB123±37%
Ornith-1.0-35B-FP8-BLOCK-MTPMoEQ8_035.5B35.21 GiB0.70 GiB36.91 GiB0.29 GiB123±37%
fable-coder-35B-A3BMoEQ8_036.0B35.21 GiB0.70 GiB36.91 GiB0.29 GiB123±37%
PINQWEN-3.6-35B-CLEAN-BF16MoEQ8_036.0B35.21 GiB0.70 GiB36.91 GiB0.29 GiB123±37%
UniMath-35B-A3BMoEQ8_036.0B35.21 GiB0.70 GiB36.91 GiB0.29 GiB123±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?
2001 of 2118 indexed open-weight models fit a A100 40GB at 131,072 context with q4_0 KV cache, the largest being NSFW_13B_sft at Q4_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.