Piccolo on Jetson and Raspberry Pi

Research by Anamika Rawat and Tulsi Patel, supervised by Dushyant Puri, through Conestoga’s SMART Centre. This work is part of SimpleDirect®’s collaboration with the Conestoga Centre for Commercialization. I’m interpreting their results here; I did not conduct or independently rerun these tests.

A $250 item, a 20 percent discount, then 13 percent tax. The price is $226. In Conestoga’s two recorded runs of this question, neither setup supplied it.

Jetson produced JSON resembling a calculate_price_with_discount_and_tax call, with correct arguments but no price. Raspberry Pi declined the calculation. Both requests were marked completed, with no truncation or runtime error. [1, slide 26]

If someone asks an application for a price, check that it actually returns the price. A speed measurement alone would have missed the problem.

Four prompts on two boards

The team used Vinci Piccolo 1.0, our roughly four-billion-parameter chat model, built from Qwen3.5-4B. We released it on June 29, 2026; the model card and download are available under Apache-2.0.

Both boards ran the Q4_K_M version of Vinci-Piccolo-1.0-GGUF through llama.cpp. Jetson Orin Nano had approximately 7.4 GiB of usable memory and Ubuntu 22.04.5. Raspberry Pi 5 had about 15 GiB and Debian 13.

Jetson used CUDA GPU acceleration; Pi used its CPU. The results show no verified Hailo-8L accelerator run after an attempted installation. The software paths and hardware both differ. [1, slide 13]

The four prompts covered renewable energy, a TypeScript algorithm, photosynthesis and the price calculation. Each ran once on each board.

Jetson’s displayed generation rates were 4.2–4.7 times Pi’s. First-token times were approximately 327–375 milliseconds versus 2.30–3.66 seconds. In thinking mode, the first text can be reasoning before the answer. [1, slides 15–25]

For the renewable-energy prompt, total request times were 171.06 seconds on Jetson and 826.31 on Pi: nearly three minutes and nearly fourteen. They generated 2,387 and 2,433 tokens respectively. That matters for an interactive application. [1, slide 16]

What needs to happen after the response

The calculation is $250 × 0.80 × 1.13 = $226. Was an application expected to execute Jetson’s tool-shaped output and return that result?

The cause remains unknown. The chat template, prompt format, settings and tool integration need checking. These two responses cannot establish how reliably Piccolo handles arithmetic in other setups.

I would test the complete interaction next, including any tool execution and the answer shown to the user. Repeat the same prompts in fresh contexts and check them independently. Our verification guide gives a way to record what actually passed.

Jetson is the faster configuration to investigate from this small sample. I’d want useful answers and tolerable waiting times on the intended workload before recommending a purchase.

What the measurements leave open

Keep the exact model file, llama.cpp build, launch commands, sampling settings, output limits, power mode, cooling and starting conditions with any repeat test.

Existing RAM use prevents these readings from isolating Piccolo’s memory needs. Different temperature sensors cannot rank whole-system cooling. Pi lacks comparable power and energy measurements, so there is no energy-efficiency ranking. Four one-off prompts provide no repeat-run reliability result or broad accuracy score.

Source record

[1] SimpleDirect Vinci-Piccolo Benchmarking.pptx, credited to Conestoga SMART Centre: configuration on slide 13; prompts on 15, 18, 21 and 24; metrics on 16, 19, 22 and 25; output and monitoring screenshots on 17, 20, 23 and 26. The supplied SimpleDirect Vinci-Piccolo Benchmarking (1).pptx repeats the substantive edge observations; it is not an independent replication.

Continue with the Mac comparison and data downloads, phone test and collaboration overview.

Research credit: Anamika Rawat and Tulsi Patel, supervised by Dushyant Puri. Benchmarking through Conestoga’s SMART Centre, within SimpleDirect®’s collaboration with the Conestoga Centre for Commercialization.