Best AI Models for NVIDIA Jetson AGX Orin 32GB
32 GB unified − 3.5 GB OS overhead = 28.5 GB available for AI models
32 GB positions this hardware in the professional tier for local AI. Most popular open-source models run comfortably, and even large 70B parameter models are accessible at lower quantization levels.
This memory amount is a sweet spot for enthusiasts and professionals. You can run 13B–30B models like DeepSeek R1 Distill at Q5 or Q6 quality with smooth token generation, and 7B models at near-lossless precision. The 70B class of models (Llama 3 70B, Qwen 72B) becomes possible at Q2–Q3 quantization, though with some quality trade-off. For day-to-day use with coding assistants, chat models, and reasoning tasks, this tier delivers an excellent experience.
Runs Well
- 7B–13B models at Q6–Q8 quality
- 14B–30B models at Q4–Q5 quality
- Small models (3B–7B) at FP16 precision
- Vision-language models at good quality
Challenging
- 70B models only at Q2–Q3 (noticeable quality loss)
- Large context windows with 30B+ models
What LLMs Can NVIDIA Jetson AGX Orin 32GB Run?
254 models · 46 excellent · 88 good
Showing compatibility for NVIDIA Jetson AGX Orin 32GB
| Model | Quant | VRAM | Speed | Context | Status | Grade |
|---|---|---|---|---|---|---|
Q4_K_M·48.9 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 16.1 GB | 48.9 t/s | 262K | RUNS WELL | A79 |
Q4_K_M·60.1 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 19.3 GB | 60.1 t/s | 262K | RUNS WELL | A84 |
Q4_K_M·49.7 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 21.4 GB | 49.7 t/s | 262K | RUNS WELL | A79 |
Q4_K_M·55.4 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 21.2 GB | 55.4 t/s | 262K | RUNS WELL | A82 |
Q4_K_M·7.6 t/s tok/s·262K ctx·DECENT | Q4_K_M | 17.4 GB | 7.6 t/s | 262K | DECENT | B45 |
Q4_K_M·7.3 t/s tok/s·131K ctx·DECENT | Q4_K_M | 18.3 GB | 7.3 t/s | 131K | DECENT | B45 |
Q4_K_M·55.8 t/s tok/s·500K ctx·RUNS WELL | Q4_K_M | 18.7 GB | 55.8 t/s | 500K | RUNS WELL | A82 |
Q4_K_M·48.9 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 16.1 GB | 48.9 t/s | 262K | RUNS WELL | A79 |
Q4_K_M·41.7 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 21.9 GB | 41.7 t/s | 262K | RUNS WELL | A76 |
Q4_K_M·40.7 t/s tok/s·203K ctx·RUNS WELL | Q4_K_M | 19.8 GB | 40.7 t/s | 203K | RUNS WELL | A75 |
Q4_K_M·153.4 t/s tok/s·128K ctx·RUNS GREAT | Q4_K_M | 5.5 GB | 153.4 t/s | 128K | RUNS GREAT | S97 |
Q4_K_M·49.7 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 21.4 GB | 49.7 t/s | 262K | RUNS WELL | A79 |
Q4_K_M·55.3 t/s tok/s·262K ctx·RUNS WELL | Q4_K_M | 18.7 GB | 55.3 t/s | 262K | RUNS WELL | A82 |
Q4_K_M·119.9 t/s tok/s·131K ctx·RUNS GREAT | Q4_K_M | 5.3 GB | 119.9 t/s | 131K | RUNS GREAT | S95 |
Q4_K_M·76.4 t/s tok/s·131K ctx·RUNS GREAT | Q4_K_M | 7.7 GB | 76.4 t/s | 131K | RUNS GREAT | S87 |
Q4_K_M·57.1 t/s tok/s·RUNS WELL | Q4_K_M | 23.3 GB | 57.1 t/s | — | RUNS WELL | A82 |
NVIDIA Jetson AGX Orin 32GB Specifications
- Brand
- NVIDIA
- Chip
- Orin
- Type
- AI Box
- Unified Memory
- 32 GB
- Memory Bandwidth
- 204.8 GB/s
- GPU Cores
- 1792
- CPU Cores
- 8
- Neural Engine
- 200.0 TOPS
- Architecture
- Ampere
- Memory Type
- LPDDR5
- TDP
- 40 W
- Release Date
- 2022-03-22
Get Started
Prompt Processing
Estimated for Mellum2 12B A2.5B Base, the compute-bound phase that reads your prompt before the first reply token appears.
Short chat
206 ms
512 tok prompt
Long chat
1.7 s
4,096 tok prompt
Document / codebase
13.2 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 2026-09-21 — see calibration basis →
Devices to Consider
Similar devices and upgrades with more memory or higher bandwidth
Frequently Asked Questions
- Can NVIDIA Jetson AGX Orin 32GB run Gemma 4 26B A4B IT?
Yes, the NVIDIA Jetson AGX Orin 32GB with 32 GB unified memory can run Gemma 4 26B A4B IT, NVIDIA Nemotron 3.5 Lightning 30B A3B BF16, Ornith 1.0 35B, and 2382 other models. 721 models achieve excellent performance, and 768 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 Jetson AGX Orin 32GB?
The NVIDIA Jetson AGX Orin 32GB has 32 GB unified memory. After macOS reserves ~3.5 GB for the operating system, approximately 28.5 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 Jetson AGX Orin 32GB good for AI?
With 32 GB unified memory and 204.8 GB/s bandwidth, the NVIDIA Jetson AGX Orin 32GB is very good for running local AI models. It supports 1489 models at good quality or better. This is a strong configuration for AI — 7B models run at maximum quality, and you can comfortably handle 14B models like DeepSeek R1 Distill 14B and larger.
- What's the best model for NVIDIA Jetson AGX Orin 32GB?
The top-rated models for the NVIDIA Jetson AGX Orin 32GB are Gemma 4 26B A4B IT, NVIDIA Nemotron 3.5 Lightning 30B A3B BF16, Ornith 1.0 35B. For general chat, instruction-tuned 7B models give the best speed-to-quality ratio. For coding or reasoning, a 14B model at Q4_K_M is a sweet spot.
- How fast is NVIDIA Jetson AGX Orin 32GB for AI inference?
With 204.8 GB/s memory bandwidth, the NVIDIA Jetson AGX Orin 32GB achieves approximately 32 tok/s on a 7B model at Q4_K_M — that's comfortable for real-time chat. A 14B model runs at ~16 tok/s. Apple Silicon achieves high efficiency (~70%) thanks to unified memory — there's no PCIe bottleneck between CPU and GPU.
tok/s = (204.8 GB/s ÷ model GB) × efficiency
Apple Silicon achieves ~70% bandwidth efficiency thanks to unified memory and Metal acceleration.
Estimated speed on NVIDIA Jetson AGX Orin 32GB
~49 tok/s~60 tok/s~50 tok/s~55 tok/sReal-world results typically within ±20%.
- Can I run AI offline on NVIDIA Jetson AGX Orin 32GB?
Yes — once you download a model, it runs entirely on the NVIDIA Jetson AGX Orin 32GB 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.