Local AI
Inside the Jetson Orin Nano Super, the computer behind Base Router
The NVIDIA board behind Base Router, our /try demo and our fleet: its specs, price history, how fast it runs language models and why 8 GB means small models.
One small NVIDIA computer sits behind three things on offlinebase.com. Base Router, the always-on AI box we are designing, runs on a Jetson Orin Nano Super. The live chat at /try (opens in a new tab) answers from a Jetson Orin Nano Super we run. The live fleet (opens in a new tab) page shows real traffic served by our own rack of Jetson Orin Nano boards.
What is on the board: 67 TOPS, 8 GB, 7 to 25 W
NVIDIA sells the board as the Jetson Orin Nano Super Developer Kit (opens in a new tab). These are NVIDIA's published specs:
Part | Jetson Orin Nano Super Developer Kit |
|---|---|
AI performance | 67 INT8 TOPS (sparse), 33 TOPS (dense) |
GPU | NVIDIA Ampere, 1,024 CUDA cores, 32 tensor cores |
CPU | 6-core Arm Cortex-A78AE, up to 1.7 GHz |
Memory | 8 GB 128-bit LPDDR5, 102 GB/s |
Storage | SD card slot and external NVMe |
Power | Modes from 7 to 25 W |
A few of those lines need translating.
TOPS means trillions of operations per second, a count of raw math. The 67 figure assumes sparsity, a shortcut that only helps models prepared for it. The dense figure, 33, is the fairer one for an ordinary model file.
The GPU runs the model's math when it writes a reply. CUDA cores handle general parallel math, and tensor cores are built for the matrix math that language models run on.
The memory is the line we watch most. The CPU and the GPU share the same 8 GB, as NVIDIA's CUDA notes for Jetson (opens in a new tab) describe. The operating system, the model and the running conversation all live in that one pool. To write each new token, the board reads most of the model's weights out of memory, so the 102 GB/s of bandwidth puts a ceiling on chat speed.
What "Super" changed, and what it costs
NVIDIA launched the Super kit on December 17, 2024 (opens in a new tab) at $249, down from $499 for the original Jetson Orin Nano Developer Kit. The gains came from higher clock speeds and a bigger power budget:
- sparse AI performance rose from 40 to 67 TOPS
- the CPU went from 1.5 to 1.7 GHz
- memory bandwidth rose from 68 to 102 GB/s
- two new power modes joined the top of the range: a 25 W mode and MAXN SUPER, which NVIDIA describes as uncapped
NVIDIA put the result at up to 1.7x higher generative-AI performance. Because the boost comes from software settings, owners of older kits get it through an update, JetPack 6.2 (opens in a new tab).
The $249 price has since gone up. NVIDIA's own Jetson FAQ (opens in a new tab) listed the Super Developer Kit at $399 when we checked on October 14, 2026. If you are pricing one for a project, that page has NVIDIA's current figure.
How fast it runs language models
Chat speed is measured in tokens per second. A token is a piece of a word, so a higher number means faster typing.
- NVIDIA's figure: Llama 3.2 3B at about 43 tokens per second on the Super kit, up from 27.7 on the original kit (NVIDIA's test, 4-bit weights on its MLC software).
- Our boards: Gemma 4 E2B answers at about 25 to 29 tokens per second. It is Google's open-weight model, and we run it as a 4-bit Q4_0 file. It is the local model behind Gemma Gemini (opens in a new tab), our free chat that picks between a Jetson and the cloud. The Gemma Gemini post explains how that choice works.
- The /try demo: Liquid LFM2.5 1.2B, a much smaller model, answers at about 38 tokens per second.
These come from different models and different software, and only the middle one is our own measurement, so read them as a rough spread for one kind of board. Smaller files generally run faster, because there are fewer bytes to read for every token.
Why 8 GB means small models
Memory decides which models fit at all. Here are three 4-bit file sizes from our memory guide:
Model | 4-bit file | On an 8 GB board |
|---|---|---|
Gemma 4 E2B | 3 GB | Fits, with room to spare |
Qwen 3.5 9B | 6 GB | Fits, but tightly: about 2 GB is left for the operating system and the conversation |
Gemma 4 12B | 8.1 GB | Too big: the estimated file alone is larger than the board's memory |
On a laptop, our configurator asks for 16 GB of RAM for the same Qwen 3.5 9B file.
So the board stays in small-model territory. Base Router's speed targets, as listed on offlinebase.com (opens in a new tab), sit at the top of that range: about 15 tokens per second for IBM Granite 4.1 8B and about 12 for Qwen 3.5 9B. Real speed depends on the model, the length of the task and what else the box is doing.
The Base Router preview covers the rest of the design.
Why 7 to 25 W suits a box that stays on
The power modes are budgets. A lower mode runs slower clocks and draws less. The highest fixed budget is 25 W, and NVIDIA also offers MAXN SUPER, which it calls an uncapped mode. That range matters for a box meant to stay on all day and answer anyone in the house at any hour.
Some plain arithmetic: a board held at 25 W around the clock would use 0.6 kWh a day. A box that waits for questions spends much of its time idle, and idle draw is far lower.
At 21:45 UTC on October 14, 2026, with no chats in flight, the four boards in our fleet that were online drew 12.1 W between them. Two weeks of real traffic on our Jetson fleet has the rest of that reading.
Our site is plain that Base Router runs warm, so we plan to give it active cooling, a small quiet fan, to keep it working through longer sessions. The site also puts its 7 to 25 W at about the draw of a lightbulb. Our energy post compares this with published data-center figures and explains where those comparisons fall short.
Where the board stops
Base Router is meant for everyday assistant work and for work on your own documents. The largest cloud systems still answer the hardest questions better, so Base Router is planned to ask your permission before it uses outside help. Base Router is in development and cannot be ordered yet.
See one working
- Chat with the board at /try (opens in a new tab). Replies stream at the pace the hardware manages, and the chats aren't stored.
- Watch /fleet (opens in a new tab) update every few seconds with tokens served and power drawn.
- To build on the boards yourself, Two weeks of real traffic on our Jetson fleet explains the free keys.