Skip to content
Run LLM Locally

GPU · NVIDIA

H100 PCIe for local LLMs: which models run?

With 80 GB of VRAM and 2000 GB/s of bandwidth, the H100 PCIe runs models up to about gpt-oss-120b entirely on the GPU.

  • VRAM80 GB HBM2e
  • Bandwidth2,000 GB/s
  • Board power350 W
  • ArchitectureHopper
ShareFacebookXWhatsAppLinkedInEmail

Specifications

VRAM
80 GB
Memory type
HBM2e
Memory bus
5120-bit
Bandwidth
2,000 GB/s
Board power
350 W
Backends
CUDA
Interface
PCIe 5.0 x16
Released
March 2023
Segment
Data center
FP16 compute
Not available

Official specData updated 24 days agoSource: NVIDIA specifications

What the H100 PCIe runs

Popular models on H100 PCIe + 64 GB RAM · 8K context
gpt-oss-120bMXFP4 · 8K~63.2 GB · 175–488 tok/s estimateExcellent
Qwen2.5 72B InstructQ6_K · 8K~62.1 GB · 23–31 tok/s estimateExcellent
Llama 3.3 70B InstructQ6_K · 8K~60.1 GB · 23–32 tok/s estimateExcellent
DeepSeek R1 Distill Llama 70BQ6_K · 8K~60.1 GB · 23–32 tok/s estimateExcellent
Mixtral 8x7B InstructQ8_0 · 8K~50.2 GB · 72–102 tok/s estimateExcellent
Qwen3 32BQ8_0 · 8K~37 GB · 37–51 tok/s estimateExcellent
DeepSeek R1 Distill Qwen 32BQ8_0 · 8K~37 GB · 37–51 tok/s estimateExcellent
Qwen2.5 32B InstructQ8_0 · 8K~37 GB · 37–51 tok/s estimateExcellent
Qwen2.5 Coder 32B InstructQ8_0 · 8K~37 GB · 37–51 tok/s estimateExcellent
QwQ 32BQ8_0 · 8K~37 GB · 37–51 tok/s estimateExcellent
Qwen3 30B-A3BQ8_0 · 8K~33.4 GB · 136–235 tok/s estimateExcellent
Qwen3 Coder 30B-A3B InstructQ8_0 · 8K~33.4 GB · 177–492 tok/s estimateExcellent
Gemma 3 27B ITQ8_0 · 8K~30.2 GB · 44–62 tok/s estimateExcellent
Gemma 2 27B ITQ8_0 · 8K~31.6 GB · 45–62 tok/s estimateExcellent
Mistral Small 3.2 24B InstructQ8_0 · 8K~26.6 GB · 53–72 tok/s estimateExcellent
gpt-oss-20bMXFP4 · 8K~13 GB · 252–695 tok/s estimateExcellent
Qwen3 14BQ8_0 · 8K~17.5 GB · 77–110 tok/s estimateExcellent
DeepSeek R1 Distill Qwen 14BQ8_0 · 8K~17.8 GB · 74–107 tok/s estimateExcellent
Qwen2.5 14B InstructQ8_0 · 8K~17.8 GB · 74–107 tok/s estimateExcellent
Phi-4 14BQ8_0 · 8K~17.6 GB · 77–110 tok/s estimateExcellent
Mistral Nemo 12B InstructQ8_0 · 8K~14.9 GB · 89–129 tok/s estimateExcellent
Gemma 3 12B ITQ8_0 · 8K~14.1 GB · 89–133 tok/s estimateExcellent
Gemma 2 9B ITQ8_0 · 8K~12.7 GB · 105–158 tok/s estimateExcellent
Qwen3 8BQ8_0 · 8K~10.5 GB · 122–184 tok/s estimateExcellent
Llama 3.1 8B InstructQ8_0 · 8K~10.2 GB · 128–191 tok/s estimateExcellent
DeepSeek R1 Distill Llama 8BQ8_0 · 8K~10.2 GB · 128–191 tok/s estimateExcellent
Qwen2.5 7B InstructQ8_0 · 8K~9.2 GB · 141–209 tok/s estimateExcellent
Mistral 7B Instruct v0.3Q8_0 · 8K~9.2 GB · 138–208 tok/s estimateExcellent
Gemma 3 4B ITQ8_0 · 8K~5.4 GB · 208–351 tok/s estimateExcellent
Qwen3 4BQ8_0 · 8K~6.2 GB · 188–313 tok/s estimateExcellent
Phi-4 Mini Instruct 3.8BQ8_0 · 8K~6 GB · 204–336 tok/s estimateExcellent
Llama 3.2 3B InstructQ8_0 · 8K~5.1 GB · 238–395 tok/s estimateExcellent
Qwen3 1.7BQ8_0 · 8K~3.7 GB · 315–579 tok/s estimateExcellent
Llama 3.2 1B InstructQ8_0 · 8K~2.5 GB · 526–946 tok/s estimateExcellent
Qwen3 0.6BQ8_0 · 8K~2.5 GB · 417–892 tok/s estimateExcellent
Llama 4 Scout 17B-16E InstructQ4_K_M · 8K~68.6 GB · 76–115 tok/s estimateCompatible
Qwen3 235B-A22BQ3_K_M · 8K~112 GB · 5–10 tok/s estimateCPU/GPU offload
DeepSeek R1 (671B)Q8_0 · 8K~699 GB · No dataIncompatible

For each model, the highest-precision quantization that runs entirely in fast memory; otherwise the one with the best result. Speeds are ranges; "estimate" unless a benchmark exists for this exact hardware.

EstimateData updated 24 days agoSource: Run LLM Locally engine

Limits in plain words

80 GB of VRAM holds the weights and KV cache of models up to about gpt-oss-120b at Q4_K_M. Bigger models spill into system RAM and slow down sharply: generation speed is then limited by RAM bandwidth, not by the GPU.

Frequently asked questions

Which LLMs can the H100 PCIe run?
Entirely on the GPU: gpt-oss-120b (MXFP4), Llama 4 Scout 17B-16E Instruct (Q4_K_M), Qwen2.5 72B Instruct (Q6_K), Llama 3.3 70B Instruct (Q6_K), DeepSeek R1 Distill Llama 70B (Q6_K), Mixtral 8x7B Instruct (Q8_0), Qwen3 32B (Q8_0), DeepSeek R1 Distill Qwen 32B (Q8_0). Larger models need CPU/GPU offload.
Can the H100 PCIe run a 70B model?
Excellent: a 70B model at Q4_K_M needs about 45.1 GB; the H100 PCIe has 80 GB.
How is the speed estimated?
Token generation reads every active weight once per token, so it is limited by memory bandwidth (2000 GB/s here). We apply an efficiency range per backend and label the result as an estimate until real benchmarks exist.