NVIDIA · datacenter

Tesla V100 32GB

Tesla V100 32GB has 32 GB of VRAM at 900 GB/s — about 29.76 GiB usable after driver and compositor overhead. 1954 of 2118 indexed models fit at 32K context with f16 KV.

Spec sheet· bandwidth, theoreticalFrom the file· fit from summed bytesPredicted· speed
Memory
32 GB
HBM2
Bandwidth
900 GB/s
4096-bit bus
Tensor FP16
125 TF
dense
TDP
300 W
Context4K8K16K32K64K128K
KV cachef16q8_0q4_0quantizing the KV cache is a ~2× lever on the dominant term at long context
text 1673vision language 177video 16embedding 26image 2audio tts 21audio asr 39

What fits at 32K context

largest quantization that fits, per model · 1954 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
grug-27bQ8_027.4B26.70 GiB2.00 GiB29.76 GiB0.00 GiB18±22%
Carnice-V2-27bQ8_027.4B26.70 GiB2.00 GiB29.76 GiB0.00 GiB18±22%
Fara1.5-27BQ8_027.4B26.70 GiB2.00 GiB29.76 GiB0.00 GiB18±22%
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.76 GiB0.00 GiB18±22%
xLAM-8x7b-rMoEIQ4_NL46.7B24.69 GiB4.00 GiB29.73 GiB0.03 GiB26±37%
Apertus-70B-Instruct-2509UD-IQ2_XXS70.6B18.54 GiB10.00 GiB29.72 GiB0.04 GiB18±22%
Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTPIQ3_M27.8B26.65 GiB2.00 GiB29.71 GiB0.05 GiB18±22%
Qwen3.5-27B-HERETIC-Polaris-Advanced-Thinking-Alpha-uncensoredQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.5-27B-Deckard-PKD-Heretic-Uncensored-ThinkingQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.6-27B-Heretic2-ThinkingQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.6-27B-abliteratedQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Webcoda-AI-27BQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
KoQweopus-3.5-27B-experimentalQ8_027.8B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Huihui-Qwen3.6-27B-abliteratedQ8_027.8B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.5-27B-hereticQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Huihui-Qwen3.5-27B-abliteratedQ8_027.8B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.5-Queen-27BQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.5-27B-abliteratedQ8_026.9B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
ThinkingCap-Qwen3.6-27B-hereticQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
MusaCoder-27BQ8_0—26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen-Image-BenchQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.5-27B-uncensored-heretic-v1Q8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Bonsai-27B-unpackedQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Ternary-Bonsai-27B-unpackedQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2-hereticQ8_027.4B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Darwin-28B-REASONQ8_026.9B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Huihui-Qwen3.5-27B-Claude-4.6-Opus-abliteratedQ8_027.8B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.5-27B-Claude-4.6-Opus-Reasoning-DistilledQ8_027.8B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.5-27BQ8_027.8B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Qwen3.5-27B-WebNovel-Writer-zhQ8_026.9B26.63 GiB2.00 GiB29.69 GiB0.07 GiB18±22%
Orca-2-13b-Alpaca-UncensoredI1-IQ2_XS13.0B3.62 GiB25.00 GiB29.67 GiB0.09 GiB18±22%
WizardLM-13B-UncensoredI1-IQ2_XS13.0B3.62 GiB25.00 GiB29.67 GiB0.09 GiB18±22%
WizardCoder-Python-13B-V1.0I1-IQ2_XS13.0B3.62 GiB25.00 GiB29.67 GiB0.09 GiB18±22%
Guanaco-13B-UncensoredI1-IQ2_XS13.0B3.62 GiB25.00 GiB29.67 GiB0.09 GiB18±22%
Open_Gpt4_8x7B_v0.1MoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
dolphin-2.6-mixtral-8x7bMoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
dolphin-2.5-mixtral-8x7bMoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
dolphin-2.7-mixtral-8x7bMoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
Nous-Hermes-2-Mixtral-8x7B-DPOMoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
Mixtral-8x7B-v0.1MoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
Mixtral-8x7B-Instruct-v0.1MoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
Mixtral-8x7B-MoE-RP-StoryMoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
Noromaid-v0.4-Mixtral-Instruct-8x7b-ZlossMoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
Open_Gpt4_8x7B_v0.2MoEQ4_046.7B24.63 GiB4.00 GiB29.66 GiB0.10 GiB26±37%
Qwen3.6-28BMoEQ8_028.2B28.00 GiB0.63 GiB29.63 GiB0.13 GiB84±37%
Qwen2.5-Coder-14B-InstructQ6_K14.8B22.58 GiB6.00 GiB29.63 GiB0.13 GiB18±22%
Bielik-11B-v2.3-InstructQ8_011.2B22.34 GiB6.25 GiB29.62 GiB0.14 GiB18±22%
Darwin-35B-A3B-OpusMoEQ6_K36.0B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
Aurora-Code-1MoEQ6_K34.7B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
grug-35b-v2MoEQ6_K35.1B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
grug-35bMoEQ6_K35.1B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
WorldSim-Opus-3.6-35B-A3BMoEQ6_K35.1B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
Qwen3.6-35B-A3B-AnkoMoEQ6_K35.1B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
KAT-Coder-V2.5-DevMoEQ6_K34.7B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
Ornith-1.0-35BMoEQ6_K34.7B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
Nex-N2-miniMoEQ6_K35.1B27.99 GiB0.63 GiB29.62 GiB0.14 GiB90±37%
Qwen3.5-99BMoEI1-IQ2_S99.0B27.80 GiB0.75 GiB29.58 GiB0.18 GiB82±37%
Qwen3.6-35B-A3BMoEUD-Q6_K36.0B27.95 GiB0.63 GiB29.58 GiB0.18 GiB90±37%
MathCoder2-CodeLlama-7BF166.7B12.55 GiB16.00 GiB29.58 GiB0.18 GiB18±22%
Llama-2-7b-chat-hfF166.7B12.55 GiB16.00 GiB29.58 GiB0.18 GiB18±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 Tesla V100 32GB run?
1954 of 2118 indexed open-weight models fit a Tesla V100 32GB at 32,768 context with f16 KV cache, the largest being grug-27b at Q8_0. That covers text, vision-language, image, video and speech models.
How much usable memory does a Tesla V100 32GB actually have?
Its nameplate is 32 GB, but about 29.76 GiB is available to a model once driver and compositor overhead is accounted for.
Is a Tesla V100 32GB fast for local AI?
Its memory bandwidth is 900 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.