GPU × LLMHopper · 141GB VRAM

Best LLM Models for the NVIDIA H200 141GB

Maximum VRAM for the largest open-weight models in a single GPU

The H200 141GB doubles the H100's VRAM and adds 21% more bandwidth, making it the first single GPU that can run DeepSeek R1 671B in INT4 or Llama 3.1 405B in BF16 without tensor parallelism. This dramatically simplifies serving infrastructure — no NVLink fabric, no multi-GPU coordination overhead. For organisations running 70B–405B models at scale, the H200 delivers lower latency per token than a 2×H100 setup while reducing operational complexity.

VRAM fit key: BF16 fit — runs in full precision INT4 fit — runs quantised API only — closed-weight, served via API

Recommended Models (141GB Hopper)

Common Use Cases

large model inferenceMoE modelssingle-GPU 405Breasoning models

VRAM Sizing Guide for NVIDIA H200 141GB

BF16 (full precision): requires ~2 bytes per parameter. On 141GB VRAM, the maximum model size in BF16 is approximately 72192B parameters (leaving ~10% headroom for KV cache).

INT8 quantisation: halves memory to ~1 byte/param — fits up to ~144384B parameters. Minimal quality loss on most models.

INT4 / Q4 quantisation: ~0.5 bytes/param — fits up to ~288768B parameters. Recommended for 70B+ models on this GPU. Use GPTQ or AWQ for best quality.

KV cache: grows with context length and batch size. At 128K context with batch size 8, a 70B model in INT4 needs ~12GB additional VRAM for KV cache. Factor this into your serving configuration.