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. 1945 of 2118 indexed models fit at 64K context with f16 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 1664vision language 177image 2audio asr 39audio tts 21video 16embedding 26

What fits at 64K context

largest quantization that fits, per model · 1945 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
OLMo-2-1124-7B-InstructQ4_K_M7.3B4.16 GiB32.00 GiB37.19 GiB0.01 GiB24±22%
Gemma4-Gutenberg-31BQ6_K31.3B24.89 GiB11.17 GiB37.14 GiB0.06 GiB25±22%
gemma-4-31B-itQ6_K31.3B24.89 GiB11.17 GiB37.14 GiB0.06 GiB25±22%
Gemma4-Gutenberg-31B-HereticQ6_K31.3B24.89 GiB11.17 GiB37.14 GiB0.06 GiB25±22%
Equinox-31BQ6_K31.3B24.89 GiB11.17 GiB37.14 GiB0.06 GiB25±22%
gemma-4-31B-it-SDFT-Heretic-RPQ6_K30.7B24.89 GiB11.17 GiB37.14 GiB0.06 GiB25±22%
GLM-Z1-32B-0414Q8_032.6B32.24 GiB3.81 GiB37.14 GiB0.06 GiB25±22%
GLM-4-32B-0414-Korean-CultureQ8_032.6B32.24 GiB3.81 GiB37.14 GiB0.06 GiB25±22%
GLM-4-32B-0414Q8_032.6B32.24 GiB3.81 GiB37.14 GiB0.06 GiB25±22%
GLM-Z1-32B-0414-uncensored-heretic-v2Q8_032.6B32.24 GiB3.81 GiB37.14 GiB0.06 GiB25±22%
Qwen3-53B-A3B-2507-THINKING-TOTAL-RECALL-v2-MASTER-CODERMoEI1-Q3_K_L53.0B25.64 GiB10.50 GiB37.13 GiB0.07 GiB36±37%
DeepSeek-R1-Distill-Llama-70BUD-IQ1_M70.6B15.99 GiB20.00 GiB37.12 GiB0.08 GiB25±22%
Hermes-4-70BUD-IQ1_M70.6B15.97 GiB20.00 GiB37.09 GiB0.11 GiB25±22%
Llama-3.3-70B-InstructUD-IQ1_M70.6B15.97 GiB20.00 GiB37.09 GiB0.11 GiB25±22%
Huihui-GLM-4.7-Flash-abliterated-57BMoEI1-IQ4_XS57.3B27.68 GiB8.37 GiB37.09 GiB0.11 GiB43±37%
Trinity-2-Codestral-22B-v0.2Q8_022.2B22.02 GiB14.00 GiB37.08 GiB0.12 GiB25±22%
Mistral-Small-Drummer-22BQ8_022.2B22.02 GiB14.00 GiB37.08 GiB0.12 GiB25±22%
Mistral-Small-Instruct-2409Q8_022.2B22.02 GiB14.00 GiB37.08 GiB0.12 GiB25±22%
Cydonia-v1.3-Magnum-v4-22BQ8_022.2B22.02 GiB14.00 GiB37.08 GiB0.12 GiB25±22%
Mistral-Small-22B-ArliAI-RPMax-v1.1Q8_022.2B22.02 GiB14.00 GiB37.08 GiB0.12 GiB25±22%
magnum-v4-22bQ8_022.2B22.02 GiB14.00 GiB37.08 GiB0.12 GiB25±22%
Codestral-22B-v0.1-hfQ8_022.2B22.02 GiB14.00 GiB37.08 GiB0.12 GiB25±22%
Codestral-22B-v0.1Q8_022.2B22.02 GiB14.00 GiB37.08 GiB0.12 GiB25±22%
dolphin-2.9.1-mixtral-1x22bMoEQ8_022.2B22.01 GiB14.00 GiB37.07 GiB0.13 GiB14±37%
ALIA-40b-fc-2606I1-Q4_140.4B23.93 GiB12.00 GiB37.04 GiB0.16 GiB25±22%
ALIA-40b-instruct-2606I1-Q4_140.4B23.93 GiB12.00 GiB37.04 GiB0.16 GiB25±22%
deepseek-coder-6.7B-kexerQ4_16.7B3.95 GiB32.00 GiB36.97 GiB0.23 GiB25±22%
WizardLM-7B-UncensoredI1-Q4_16.7B3.95 GiB32.00 GiB36.97 GiB0.23 GiB25±22%
Llama-2-7B-32K-InstructI1-Q4_16.7B3.95 GiB32.00 GiB36.97 GiB0.23 GiB25±22%
EuroLLM-22B-Instruct-2512Q8_022.6B22.41 GiB13.50 GiB36.97 GiB0.23 GiB25±22%
SambaLingo-Japanese-ChatI1-Q4_K_M6.9B3.94 GiB32.00 GiB36.96 GiB0.24 GiB25±22%
Magistral-Small-2509-VisionQ6_K_L24.0B25.83 GiB10.00 GiB36.94 GiB0.26 GiB25±22%
MN-GRAND-23.5B-Gutenberg-UNCENSORED-V2-GLM4.7-ThinkingQ5_K_M23.4B15.64 GiB20.25 GiB36.94 GiB0.26 GiB25±22%
Qwen3.5-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-ThinkingI1-Q6_K39.5B29.87 GiB6.00 GiB36.94 GiB0.26 GiB25±22%
Qwen3.5-40B-RoughHouse-Claude-4.6-Opus-Polar-Deckard-Uncensored-Heretic-ThinkingI1-Q6_K39.5B29.87 GiB6.00 GiB36.94 GiB0.26 GiB25±22%
Qwen3.5-40B-Claude-4.5-Opus-High-Reasoning-Thinking-uncensored-hereticQ6_K39.5B29.87 GiB6.00 GiB36.94 GiB0.26 GiB25±22%
Apertus-70B-Instruct-2509IQ1_M70.6B15.74 GiB20.00 GiB36.92 GiB0.28 GiB25±22%
Mistral-Small-4-119B-2603MoEIQ2_XS119B34.48 GiB1.41 GiB36.92 GiB0.28 GiB107±37%
MathCoder2-CodeLlama-7BQ4_K_L6.7B3.89 GiB32.00 GiB36.92 GiB0.28 GiB25±22%
Qwen3-16B-A3BMoEBF1616.0B29.88 GiB6.00 GiB36.87 GiB0.33 GiB44±37%
Qwythos-9B-Claude-Mythos-5-1MBF169.4B33.83 GiB2.00 GiB36.87 GiB0.33 GiB25±22%
Qwythos-9B-v2BF169.7B33.83 GiB2.00 GiB36.87 GiB0.33 GiB25±22%
llm-jp-4-32b-a3b-thinkingMoEQ8_032.1B31.84 GiB4.00 GiB36.84 GiB0.36 GiB63±37%
Qwen3.6-35B-A3B-REAM-192-hereticMoEQ5_K_M27.0B34.58 GiB1.25 GiB36.84 GiB0.36 GiB103±37%
deepseek-coder-6.7b-instructQ4_K_M6.7B3.80 GiB32.00 GiB36.83 GiB0.37 GiB25±22%
deepseek-coder-6.7b-baseQ4_K_M6.7B3.80 GiB32.00 GiB36.83 GiB0.37 GiB25±22%
Magicoder-S-DS-6.7BI1-Q4_K_M6.7B3.80 GiB32.00 GiB36.83 GiB0.37 GiB25±22%
internlm3-8b-instructF328.8B32.80 GiB3.00 GiB36.83 GiB0.37 GiB25±22%
CodeLlama-7b-instruct-hfQ4_K_M6.7B3.80 GiB32.00 GiB36.83 GiB0.37 GiB25±22%
CodeLlama-7b-hfQ4_K_M6.7B3.80 GiB32.00 GiB36.83 GiB0.37 GiB25±22%
Luna-AI-Llama2-UncensoredI1-Q4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
Llama-2-7b-chat-hfQ4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
Swallow-7b-NVE-instruct-hfI1-Q4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
llava-v1.5-7bQ4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
CodeLlama-7b-python-hfQ4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
Wizard-Vicuna-7B-UncensoredQ4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
llama2_7b_chat_uncensoredQ4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
WizardLM-7B-V1.0-UncensoredQ4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
Llama-2-7b-hfQ4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 GiB25±22%
pygmalion-2-7bQ4_K_M6.7B3.80 GiB32.00 GiB36.82 GiB0.38 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?
1945 of 2118 indexed open-weight models fit a A100 40GB at 65,536 context with f16 KV cache, the largest being OLMo-2-1124-7B-Instruct 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.