The Jetson Orin Nano Super costs $399 now. It launched at $249 in December 2024, and NVIDIA quietly moved it up in July 2026 with no announcement. That price change reopened the obvious question: what is the cheapest way to run a local LLM on a small, low-power box at the edge?
Three boards get named every time. The Jetson Orin Nano Super, the Raspberry Pi AI HAT+ 2 on a Pi 5, and the same Hailo-10H chip sold as an M.2 card or USB stick for other machines. On paper the Hailo parts look like a cheap shortcut to generative AI. The published benchmarks say something different, and the fastest option per dollar is not the one with the AI label on the box. Prices and benchmark figures below were checked against NVIDIA, Raspberry Pi, Hailo and retailer listings on October 12, 2026.
The short answer
- Fastest tokens per dollar: the Jetson Orin Nano Super. It runs ordinary GGUF and Hugging Face models on a CUDA GPU with 102 GB/s of memory bandwidth, and published numbers put 3B-class models at 22 to 43 tokens per second depending on the model and runtime.
- Cheapest box that runs real models: a Raspberry Pi 5 with no accelerator. The CPU beats the AI HAT+ 2 on decode speed in every independent test we found, and the 16GB board can hold models the HAT cannot load at all.
- Lowest power for vision plus a small LLM: the AI HAT+ 2. It wins on peak power under load, not on speed.
- Hailo-10H for a PC you already own: the ASUS UGen300, as a USB stick or an M.2 card. Same chip, same model list, and on an x86 host the CPU or iGPU is even harder to beat.
If you only want one recommendation, buy the Jetson and treat the price rise as the cost of a real GPU. The rest of this guide shows why the cheaper options are slower than they look, and when each one still makes sense.
How we compared them
This is a research comparison, not a lab test. We did not benchmark these boards ourselves. The specs come from NVIDIA, Raspberry Pi, Hailo and ASUS. The token rates come from the NVIDIA launch numbers, the published Hailo model zoo table, and independent reviews by Jeff Geerling and CNX Software. Every number below names the runtime it came from, because the same board can double its speed between Ollama and an optimized engine.
Prices are US list prices from NVIDIA and approved Raspberry Pi resellers, plus the live Amazon listing where one exists. Amazon prices for Pi hardware often run above list, so we call that out where it matters.
Specs at a glance
| Jetson Orin Nano Super | Pi 5 + AI HAT+ 2 | ASUS UGen300 (Hailo-10H) | Pi 5 16GB, CPU only | |
|---|---|---|---|---|
| AI compute | 67 TOPS INT8 (sparse), Ampere GPU | 40 TOPS INT4 | 40 TOPS INT4 | 4 Cortex-A76 cores |
| Memory for the model | 8GB LPDDR5, shared | 8GB LPDDR4X on the HAT | 8GB LPDDR4 on the module | 16GB LPDDR4X, shared |
| Memory bandwidth | 102 GB/s | Not published | Not published | Up to 17 GB/s (32-bit) |
| Runs standard GGUF files | Yes (llama.cpp, Ollama); vLLM and MLC also supported | No, Hailo HEF files only | No, Hailo HEF files only | Yes (llama.cpp, Ollama) |
| Largest LLM in practice | About 8B at 4-bit | 1.7B LLM, 2B VLM (Hailo zoo) | 1.7B LLM, 2B VLM (Hailo zoo) | 30B MoE with a tight quant |
| Power | 7 to 25W modes plus MAXN | 7.1W max, whole Pi under LLM load | 2.5W typical for the chip | 12.3W max under LLM load (measured on a Pi 5 8GB) |
| Price, Oct 2026 | About $399 | About $200 for the HAT plus the Pi | About $300 (USB); M.2 lists at about $260 but has a 4 to 5 week lead time | About $305 list |
Memory bandwidth decides how fast a model writes tokens, more than TOPS does. Every generated token reads the whole model out of RAM. The Jetson has a 128-bit LPDDR5 bus at 102 GB/s. The Pi 5 has a 32-bit bus rated up to 17 GB/s, about a sixth of that, and that gap is most of the story in the benchmarks below.
Published LLM speeds
Two tables, because no single reviewer ran all four. The first is the Pi side, from the AI HAT+ 2 review by Jeff Geerling on a Pi 5 8GB. The HAT used the compiled models from Hailo through hailo-ollama, and the CPU used stock Ollama on the same Pi.
| Model | AI HAT+ 2 (tok/s) | Pi 5 CPU (tok/s) |
|---|---|---|
| qwen2 1.5B | 8.25 | 12.47 |
| qwen2.5-instruct 1.5B | 6.86 | 11.41 |
| qwen2.5-coder 1.5B | 8.13 | 9.77 |
| deepseek-r1-distill-qwen 1.5B | 6.85 | 10.99 |
| llama3.2 3B | 2.66 | 5.55 |
The CPU won every row. CNX Software saw the same pattern on a 2GB Pi 5, and the Hailo model zoo table lists every LLM below 10 tokens per second, with Llama 3.2 1B at the top at 9.89. The 3B Llama has since been dropped from the zoo.
The second table is the Jetson. The NVIDIA numbers use the MLC engine with INT4 weights, and they are best case. The independent ClawBox runs used plain Ollama on a production Orin Nano 8GB, which is closer to what most people install first.
| Model | NVIDIA, MLC INT4 (tok/s) | Ollama, independent (tok/s) |
|---|---|---|
| Llama 3.2 1B | Not published | 44.9 |
| Gemma 2 2B | 34.97 | 25.6 |
| Llama 3.2 3B / Qwen2.5 3B | 43.07 (Llama) | 22.6 (Qwen) |
| Qwen2.5 7B | 21.75 | Not tested |
| Llama 3.1 8B | 19.14 | Not tested |
Even the slower Ollama figure, Qwen2.5 3B at 22.6, is about four times what the Pi CPU manages on Llama 3.2 3B and eight times the HAT. The Pi options are fine for a short answer and slow for anything long.
Jetson Orin Nano Super
This is NVIDIA’s own developer kit, and the only one of the four that runs CUDA.
Verdict: the only board here that runs local LLMs at a speed you would use every day. It also runs the normal toolchain: llama.cpp, Ollama, vLLM, TensorRT-LLM and MLC are all on the NVIDIA support list, so any GGUF you already have will load.
The module is a 6-core Cortex-A78AE with a 1,024-core Ampere GPU and 32 Tensor cores. The Super name is a power mode, not new silicon. The same board got a higher clock and the 102 GB/s memory rate through JetPack. The original Orin Nano kit can be brought up to the same speed with a software update. NVIDIA rates it for models up to about 8B parameters. In practice the 8GB is shared with the OS, so 3B to 4B models leave more room for context and other services.
Good fit: anyone who wants a private assistant, a coding model or a voice pipeline on a 15 to 25W box. Skip it if: you need more than 8GB, or you cannot stand a price that jumped 60% in one step. From the JetPack 7 branch there is no SD card image for this kit, so plan on flashing from a USB stick and booting an NVMe drive.
The NVIDIA price table lists the kit at $399. CNX Software covered the July repricing, which hit the whole Jetson range by up to 101%. Stock has been patchy since, so check before you plan around it. Its Amazon listing went out of stock on October 19, so the buttons here open NVIDIA’s own page for the kit, which links to NVIDIA’s partner store.
Raspberry Pi 5 with the AI HAT+ 2
The HAT is the Raspberry Pi route to a Hailo-10H, and it is built to sit on a Pi 5.
Verdict: a good vision accelerator that also runs small LLMs, slowly. The Hailo-10H has 8GB of its own LPDDR4X, so it does not eat into Pi memory. Hailo lists time to first token at 0.32 to 0.73 seconds for its 1.5B models. Neither independent review measured it against the CPU, so treat that as the vendor figure.
The limits matter more than the TOPS figure. The HAT only runs models Hailo has compiled into its HEF format. You cannot point it at a GGUF from Hugging Face. Every LLM in the zoo uses a 2,048-token context, and the biggest current LLM is Qwen3 1.7B, with 2B vision models alongside it. The Raspberry Pi launch post talked about 1 to 7 billion parameter models. The shipping model list is well short of that.
It also takes the Pi 5’s single PCIe lane. With the HAT fitted, there is no NVMe boot drive unless you add a PCIe switch board, so the OS lives on a microSD card or a USB drive.
Who should buy it: camera projects that also want a small local model for captions, commands or tool calls, especially on battery or solar. Geerling measured 7.1W peak for the whole Pi with the HAT under LLM load, against 12.3W running the model on the CPU. At idle the order flips: 4.2W with the HAT, 3.1W without. The saving only appears while the model runs. Skip it if: text generation is the main job. The Pi CPU is faster.
Price has moved twice. The HAT launched at $130 in January, went up $50 in the Raspberry Pi April price rise, and now sells for about $200 at approved resellers. The Amazon listing for the bare HAT is out of stock, and the third-party kits there cost far more, so buy it from PiShop or SparkFun.
Hailo-10H outside the Pi: ASUS UGen300
The UGen300 is how ASUS sells the Hailo-10H for machines that are not a Pi.
Verdict: the practical way to buy a Hailo-10H for a mini PC or laptop, but not a reason to buy one for LLMs. The Hailo-branded M.2 module has been hard for individuals to buy. Threads on the Hailo community forum describe empty distributor shelves and quotes in the thousands, with only a recent report of a 4GB version in stock at one distributor. ASUS packages the same chip with 8GB of LPDDR4 in two forms. The UGen300 USB connects over USB 3.1 Gen 2 Type-C. The UGen300 M.2 is a 2280 Key-M card on PCIe 3.0 x4.
Software is the same HailoRT stack and the same model zoo, on Linux or Windows. So the 1.7B LLM ceiling (2B for vision models) and the 2,048-token context come with it. We found no independent LLM benchmark of the Hailo-10H on an x86 host yet, so expect the sub-10 tokens per second. On a PC the comparison point is a much faster CPU than the one in a Pi, which makes the case for the NPU on raw speed weaker still.
Best for: an existing NVR, kiosk or industrial PC that needs offloaded vision with low CPU load, plus a small model on the side. Skip it if: you are buying it to chat with a model. Spend the money on RAM or a GPU instead. On M.2 boxes, check you are not giving up the only SSD slot.
The USB version is sold by Amazon.com at about $299, its list price. The M.2 card lists at about $260, but Newegg listed it as out of stock on October 12, and on October 19 its only Amazon listing showed a 4 to 5 week lead time at about $320 from a third-party seller. Until that clears, the USB stick is the version you can get quickly.
Raspberry Pi 5 16GB, no accelerator
Verdict: the cheapest box here that runs real, current models from the open GGUF ecosystem. It is slow, about 5 to 12 tokens per second on 1.5B to 3B models. But it loads anything llama.cpp supports, and the 16GB board fits far more than the 8GB on the HAT. Geerling ran Qwen3 30B-A3B on one with a tight quant, because a mixture-of-experts model only reads a small slice of its weights per token.
Memory prices have hit the Pi hard. The 16GB board launched at $120 and now lists at $305 after three rounds of increases. If you already own a Pi 5 8GB, start there. It runs the same 1.5B to 3B models shown above. Our llama.cpp guide covers the build, and the quantization comparison explains which GGUF quant to pick for a small board.
Who it suits: tinkerers who want to run any model, slowly, for under $400 all in. Skip it if: you want speed. Amazon has no bare 16GB board in stock right now. The closest listing is a RasTech kit with the board, a third-party 27W GaN supply, an Active Cooler, a 64GB card and a case at about $390, roughly $20 over the CPU-only build total below, while approved resellers such as PiShop sell the bare board at its $305 list price. Check the RasTech Pi 5 16GB kit on Amazon.
What a working box costs
The board price is not the build price. These totals use list prices from approved resellers in October 2026 and include the parts each board needs to boot. None includes a case.
| Build | Parts | Approx. total |
|---|---|---|
| Jetson Orin Nano Super | Dev kit with 19V supply, plus an NVMe drive (or microSD on JetPack 6) | About $440 to $480 |
| Pi 5 8GB + AI HAT+ 2 | Pi $175, HAT $200, 27W supply, Active Cooler, 64GB microSD | About $440 |
| Pi 5 2GB + AI HAT+ 2 | Pi $77.50, HAT $200, supply, cooler, microSD | About $340 |
| Pi 5 16GB, CPU only | Pi $305, supply, cooler, microSD | About $370 |
| Pi 5 8GB, CPU only | Pi $175, supply, cooler, microSD | About $240 |
| ASUS UGen300 on a PC you own | USB stick (the M.2 card has a 4 to 5 week lead time) | About $300 |
The Jetson and the Pi 5 8GB plus HAT builds are the useful comparison. For about the same money, the Jetson writes tokens three to eight times faster and runs standard GGUF models. The HAT build draws less power and keeps the Pi GPIO ecosystem. A 2GB Pi works as the host because the model lives in the memory on the HAT, which is the only way the HAT path gets cheaper.
What to check before you buy
- Bandwidth over TOPS. TOPS figures measure matrix math, mostly for vision. LLM decode speed tracks memory bandwidth and model size. A 40 TOPS NPU on a narrow memory bus loses to a GPU with a wide one.
- Model format lock-in. Hailo runs only its own compiled models. If the model you want is not in the zoo, the hardware cannot run it. Jetson and the Pi CPU take standard GGUF files.
- Context length. The 2,048-token Hailo limit rules out long documents and most RAG setups. Check the context your workload needs before you look at speed.
- Memory headroom. 8GB shared with an OS is tight. A 4-bit 3B model plus a few thousand tokens of context fits easily. An 8B model fits with little room left for anything else.
- Buy now or wait. As of October 2026, memory prices are still rising. NVIDIA repriced the whole Jetson line in July, and Raspberry Pi raised prices in February, April and October. Neither company has announced a cut. If you need the hardware this quarter, buy at list from an approved seller and compare any third-party listing against list price first. Several Amazon Pi listings run about 14% over list.
If a heavier model is the real goal, an edge board is the wrong tool. Our mini PC for local AI guide and the Mac mini vs mini PC vs GPU comparison cover the next step up. For the Pi itself, the Pi 5 kits and alternatives guide lists the boards and bundles worth buying, and Hailo vs Coral for Frigate covers the vision work the Hailo chips are actually built for. If the Pi is meant to be a home server rather than an AI box, our Raspberry Pi 5 vs N150 comparison prices it against a mini PC.
Pick by workload
Chat, coding help or a voice assistant: the Jetson Orin Nano Super, run through llama.cpp or Ollama at first and MLC later if you want the NVIDIA numbers. Any model on the cheapest hardware: a Pi 5, CPU only, with the 16GB board if you want room to grow. Cameras first and a small model second, on a tight power budget: the AI HAT+ 2. A Hailo-10H in a PC you already own: the ASUS UGen300, bought for vision. On every budget here, the LLM bottleneck is memory bandwidth. Pay for that before you pay for TOPS.
