The NVIDIA H200 NVL packs 141GB of HBM3e with 4890 GB/s of memory bandwidth, making it one of the fastest cards you can run a local model on. Here is which models fit, how fast they run, and what quantisation gives the best balance of quality and speed.
Already have it? Find the best model for it. Still choosing? See what else $35,000 buys. Or the full specifications.
Two specs decide almost everything when you run a language model locally. Memory capacity decides which models fit at all, and memory bandwidth decides how fast they write, because every generated token reads the whole model once.
With 141GB of VRAM, the NVIDIA H200 NVL holds models up to roughly 246B parameters at Q4_K_M, the most common quantisation, once the runtime has taken its own share. The context you ask for needs room on top of that, so a model near the limit runs with a short context.
At 4890 GB/s, speed falls as models grow: a model twice the size generates at roughly half the tokens per second. The estimates in the tables come from the bandwidth and each model's size, and real runs differ by 10–20% with context length and the inference backend.
For general chat and everyday use on the NVIDIA H200 NVL, start with a 14B–32B model at Q4_K_M. It is the best balance of answer quality and generation speed this card offers.
For coding, pick a model built for it (“coder” in the name). For reasoning-heavy work, reasoning models such as DeepSeek R1 distills or Qwen with thinking mode do best, at the cost of longer answers.
The NVIDIA H200 NVL specs page has the full specifications and the MSRP ($35,000).