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. 2011 of 2118 indexed models fit at 16K context with q8_0 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
video 16text 1726vision language 181image 2embedding 26audio tts 21audio asr 39

What fits at 16K context

largest quantization that fits, per model · 2011 of 2118 indexed
ModelBest quantParamsWeights●KV●Total in memory◐Headroom◐tok/s◐
Bernini-RQ8_014.3B28.71 GiB0.00 GiB29.76 GiB0.00 GiB18±22%
command-r-35b-writer-v2IQ4_XS35.0B18.02 GiB10.63 GiB29.75 GiB0.01 GiB18±22%
Phi-3.5-MoE-instructMoEKV unresolvedQ5_K_M41.9B27.68 GiB1.06 GiB29.74 GiB0.02 GiB48±37%
Gemma-4-Novelist-Eclipse-31BQ6_K_L32.7B26.60 GiB1.95 GiB29.64 GiB0.12 GiB18±22%
Gemma-4-31B-StyleTuneQ6_K_L32.7B26.60 GiB1.95 GiB29.64 GiB0.12 GiB18±22%
Salience-1.5-ProMoEQ6_K36.0B28.43 GiB0.17 GiB29.60 GiB0.16 GiB103±37%
Qwable-v1MoEQ6_K36.0B28.43 GiB0.17 GiB29.60 GiB0.16 GiB103±37%
T-SearchMoEQ6_K36.0B28.43 GiB0.17 GiB29.60 GiB0.16 GiB103±37%
MiniMax-M2.1-REAP-139B-A10BMoEI1-IQ1_S139B26.55 GiB2.06 GiB29.59 GiB0.17 GiB58±37%
m51Lab-MiniMax-M2.7-REAP-139B-A10BMoEI1-IQ1_S139B26.55 GiB2.06 GiB29.59 GiB0.17 GiB58±37%
Assistant_Pepe_70BQ2_K70.6B25.79 GiB2.66 GiB29.58 GiB0.18 GiB18±22%
Hermes-4-70BUD-IQ3_XXS70.6B25.76 GiB2.66 GiB29.54 GiB0.22 GiB18±22%
Llama-3.3-70B-InstructUD-IQ3_XXS70.6B25.76 GiB2.66 GiB29.54 GiB0.22 GiB18±22%
DeepSeek-R1-Distill-Llama-70BUD-IQ3_XXS70.6B25.76 GiB2.66 GiB29.54 GiB0.22 GiB18±22%
Apriel-1.6-15b-ThinkerBF1614.9B26.88 GiB1.59 GiB29.53 GiB0.23 GiB18±22%
Qwen3.5-88BMoEI1-Q2_K_S87.7B28.30 GiB0.20 GiB29.52 GiB0.24 GiB91±37%
Kimi-Linear-48B-A3B-InstructMoEQ4_K_L49.1B28.26 GiB0.25 GiB29.52 GiB0.24 GiB18±22%
ALIA-40b-fc-2606I1-Q5_K_M40.4B26.78 GiB1.59 GiB29.49 GiB0.27 GiB18±22%
ALIA-40b-instruct-2606I1-Q5_K_M40.4B26.78 GiB1.59 GiB29.49 GiB0.27 GiB18±22%
Hunyuan-A13B-InstructMoEQ2_K80.4B27.40 GiB1.06 GiB29.46 GiB0.30 GiB18±22%
Qwen3-Coder-NextMoEUD-IQ3_S79.7B27.65 GiB0.80 GiB29.44 GiB0.32 GiB94±37%
CalmeRys-78B-Orpo-v0.1I1-IQ2_XXS78.0B25.43 GiB2.86 GiB29.41 GiB0.35 GiB18±22%
calme-2.3-rys-78bIQ2_XXS78.0B25.43 GiB2.86 GiB29.41 GiB0.35 GiB18±22%
Qwen3.6-27B-Heretic2-Uncensored-Finetune-ThinkingQ8_027.4B27.82 GiB0.53 GiB29.41 GiB0.35 GiB18±22%
Devstral-2-123B-Instruct-2512IQ1_S125B25.33 GiB2.92 GiB29.41 GiB0.35 GiB18±22%
Mistral-Medium-3.5-128BI1-IQ1_S128B25.33 GiB2.92 GiB29.41 GiB0.35 GiB18±22%
XORTRON-NXTXPRTXXLI1-IQ1_S128B25.33 GiB2.92 GiB29.41 GiB0.35 GiB18±22%
Darwin-35B-A3B-OpusMoEQ6_K_L36.0B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
Aurora-Code-1MoEQ6_K_L34.7B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
grug-35b-v2MoEQ6_K_L35.1B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
grug-35bMoEQ6_K_L35.1B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
WorldSim-Opus-3.6-35B-A3BMoEQ6_K_L35.1B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
Qwen3.6-35B-A3B-AnkoMoEQ6_K_L35.1B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
KAT-Coder-V2.5-DevMoEQ6_K_L34.7B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
Ornith-1.0-35BMoEQ6_K_L34.7B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
Nex-N2-miniMoEQ6_K_L35.1B28.22 GiB0.17 GiB29.39 GiB0.37 GiB103±37%
Apertus-70B-Instruct-2509IQ3_XXS70.6B25.53 GiB2.66 GiB29.37 GiB0.39 GiB18±22%
Maenad-70BI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
DeepSeek-R1-Distill-Llama-70B-Uncensored-v2-Unbiased-ReasonerI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Rombos-LLM-70b-Llama-3.3I1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
L3.3-Electra-R1-70bI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
L3.3-70B-Magnum-v4-SEIQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Latxa-Llama-3.1-70B-Instruct-v2I1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Llama-3.3_70_b_uncensored_continuedI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Llama-3.3-70B-Instruct-abliteratedI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
grok-oss-Revenant-70BI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Llama-3.1-Nemotron-70B-Instruct-HFI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
L3.3-70B-Euryale-v2.3I1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Hermes-3-Llama-3.1-70BIQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Hermes-4-70B-hereticI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Llama-3.1-70BIQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Anubis-70B-v1.2IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Golem-70B-v1bI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
DeepSeek-R1-Distill-Llama-70B-abliteratedI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
DeepSeek-R1-Distill-Llama-70B-hereticI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Legion-V2.1-LLaMa-70BI1-IQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Tess-R1-Limerick-Llama-3.1-70BIQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
SEMIKONG-70BIQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
functionary-medium-v3.2KV unresolvedIQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 GiB18±22%
Llama-3.1-WhiteRabbitNeo-2-70BIQ3_XXS70.6B25.58 GiB2.66 GiB29.36 GiB0.40 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?
2011 of 2118 indexed open-weight models fit a Tesla V100 32GB at 16,384 context with q8_0 KV cache, the largest being Bernini-R 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.