Best AI Models for NVIDIA DGX H100
640 GB unified − 3.5 GB OS overhead = 636.5 GB available for AI models
With 640 GB of memory, this is a high-end configuration for local AI. You can comfortably run most open-source LLMs including large 70B parameter models at good quantization levels, making it one of the best setups for serious local AI work.
At this memory tier, nearly every popular open-source model is within reach. You can run Llama 3 70B at Q4_K_M or even Q5_K_M quantization with room to spare, handle coding assistants like DeepSeek Coder 33B at high quality, and easily run any 7B–30B model at full or near-full precision. Context windows remain generous even with larger models, so multi-turn conversations and long-document processing work smoothly.
Runs Well
- 70B models (Llama 3 70B, Qwen 72B) at Q4–Q5
- 30B models at Q6–Q8 quality
- 7B–14B models at full FP16 precision
- Vision models (LLaVA, CogVLM) without compromise
Challenging
- Mixture-of-experts models like Mixtral 8x22B at higher quants
- 120B+ models still require lower quantizations
What LLMs Can NVIDIA DGX H100 Run?
175 models · 155 excellent · 13 good
Showing compatibility for NVIDIA DGX H100
| Model | Quant | VRAM | Speed | Context | Status | Grade |
|---|---|---|---|---|---|---|
Q4_K_M·42.0 t/s tok/s·164K ctx·RUNS WELL | Q4_K_M | 414.6 GB | 42.0 t/s | 164K | RUNS WELL | A76 |
Q4_K_M·38.2 t/s tok/s·1049K ctx·RUNS WELL | Q4_K_M | 456.2 GB | 38.2 t/s | 1049K | RUNS WELL | A74 |
Q4_K_M·42.0 t/s tok/s·164K ctx·RUNS WELL | Q4_K_M | 415.1 GB | 42.0 t/s | 164K | RUNS WELL | A76 |
Q4_K_M·42.0 t/s tok/s·164K ctx·RUNS WELL | Q4_K_M | 414.6 GB | 42.0 t/s | 164K | RUNS WELL | A76 |
Q4_K_M·42.0 t/s tok/s·164K ctx·RUNS WELL | Q4_K_M | 414.6 GB | 42.0 t/s | 164K | RUNS WELL | A76 |
Q4_K_M·42.0 t/s tok/s·164K ctx·RUNS WELL | Q4_K_M | 414.6 GB | 42.0 t/s | 164K | RUNS WELL | A76 |
Q4_K_M·42.0 t/s tok/s·164K ctx·RUNS WELL | Q4_K_M | 414.6 GB | 42.0 t/s | 164K | RUNS WELL | A76 |
Q4_K_M·42.0 t/s tok/s·164K ctx·RUNS WELL | Q4_K_M | 415.1 GB | 42.0 t/s | 164K | RUNS WELL | A76 |
Q4_K_M·47.1 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 369.9 GB | 47.1 t/s | 262K | RUNS WELL | A78 |
Q4_K_M·42.0 t/s tok/s·164K ctx·RUNS WELL | Q4_K_M | 415.1 GB | 42.0 t/s | 164K | RUNS WELL | A76 |
Q4_K_M·32.6 t/s tok/s·1049K ctx·RUNS WELL | Q4_K_M | 534.0 GB | 32.6 t/s | 1049K | RUNS WELL | A68 |
Q4_K_M·60.3 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 288.6 GB | 60.3 t/s | 262K | RUNS WELL | A84 |
Q4_K_M·65.0 t/s tok/s·RUNS WELL | Q4_K_M | 267.9 GB | 65.0 t/s | — | RUNS WELL | A84 |
Q4_K_M·67.9 t/s tok/s·1049K ctx·RUNS GREAT | Q4_K_M | 256.7 GB | 67.9 t/s | 1049K | RUNS GREAT | S85 |
Q4_K_M·65.7 t/s tok/s·RUNS GREAT | Q4_K_M | 265.1 GB | 65.7 t/s | — | RUNS GREAT | S85 |
Q4_K_M·73.0 t/s tok/s·262K ctx·RUNS GREAT | Q4_K_M | 238.5 GB | 73.0 t/s | 262K | RUNS GREAT | S86 |
NVIDIA DGX H100 Specifications
- Brand
- NVIDIA
- Chip
- 8x H100 SXM5
- Type
- Server
- Unified Memory
- 640 GB
- Memory Bandwidth
- 26800.0 GB/s
- GPU Cores
- 135168
- CPU Cores
- 112
- TDP
- 10200 W
- Interconnect
- NVSwitch 4th-gen, 900 GB/s NVLink per GPU
- MSRP
- $482,000
- Release Date
- 2022-09-01
Get Started
Prompt Processing
Estimated for GLM 5, the compute-bound phase that reads your prompt before the first reply token appears.
Short chat
217 ms
512 tok prompt
Long chat
1.7 s
4,096 tok prompt
Document / codebase
13.9 s
32,768 tok prompt
Prefill is compute-bound and a different number from the decode tok/s shown elsewhere on this page — how prompt processing works →
Performance figures are estimates calibrated as of 2026-07-30 — see calibration basis →
Devices to Consider
Similar devices and upgrades with more memory or higher bandwidth
Frequently Asked Questions
- Can NVIDIA DGX H100 run DeepSeek R1?
Yes, the NVIDIA DGX H100 with 640 GB unified memory can run DeepSeek R1, GLM 5.2, DeepSeek V3.2, and 1789 other models. 1746 models achieve excellent performance, and 34 run at good quality. Apple Silicon's unified memory architecture lets the GPU access the full memory pool without copying data, making it efficient for AI workloads.
- How much memory is available for AI on NVIDIA DGX H100?
The NVIDIA DGX H100 has 640 GB unified memory. After macOS reserves ~3.5 GB for the operating system, approximately 636.5 GB is available for AI models. Unlike discrete GPUs where VRAM is separate from system RAM, Apple Silicon shares one memory pool between the CPU and GPU — this means no data copying overhead, but you share memory with macOS and open apps.
- Is NVIDIA DGX H100 good for AI?
With 640 GB unified memory and 26800.0 GB/s bandwidth, the NVIDIA DGX H100 is excellent for running local AI models. It supports 1780 models at good quality or better. This is a premium configuration — you can run large 30B+ parameter models at good quality, and most 7B models at maximum quality. Ideal for professional AI workloads.
- What's the best model for NVIDIA DGX H100?
The top-rated models for the NVIDIA DGX H100 are DeepSeek R1, GLM 5.2, DeepSeek V3.2. With this much memory, you can prioritize quality — use higher quantizations (Q5/Q6) for better output, or run larger 30B+ models for more capable reasoning.
- How fast is NVIDIA DGX H100 for AI inference?
With 26800.0 GB/s memory bandwidth, the NVIDIA DGX H100 achieves approximately 4169 tok/s on a 7B model at Q4_K_M — that's very fast, well above conversational speed. A 14B model runs at ~2084 tok/s. Apple Silicon achieves high efficiency (~70%) thanks to unified memory — there's no PCIe bottleneck between CPU and GPU.
tok/s = (26800 GB/s ÷ model GB) × efficiency
Apple Silicon achieves ~70% bandwidth efficiency thanks to unified memory and Metal acceleration.
Estimated speed on NVIDIA DGX H100
~42 tok/s~38 tok/s~42 tok/s~42 tok/sReal-world results typically within ±20%.
- Can I run AI offline on NVIDIA DGX H100?
Yes — once you download a model, it runs entirely on the NVIDIA DGX H100 without internet. Applications like Ollama and LM Studio make it straightforward to download, manage, and run models locally. All your conversations stay private on your device with zero data sent to external servers. This is one of the key advantages of local AI: complete privacy, no API costs, and no rate limits.