Explore native execution for supported edge and mobile hardware, with memory, sustained performance and deployment requirements made explicit.
Device class and launch state are separate. A device appearing in a class below does not mean it is released.
From a Raspberry Pi Zero 2 W to a Jetson AGX Orin — the same native runtime, no framework and no cloud, drawing on the device you ship.
Shown as what the engine draws on the named board — output, not a benchmark; no speed figure is attached to these images. Each was rendered by Veizik natively on the device shown, from a published checkpoint, with no desktop GPU in the loop.
What the board allows decides what you can run, before any model is chosen.
The budget is set by the board, from 8 GB on a Raspberry Pi 5 down to 512 MB on a Pi Zero 2 W. The engine streams a model through a bounded working set instead of holding all of it, and the precision tier is a flag, so the footprint is chosen per job.
Peak RAM, measured on the board: see the table below.
Multi-core inference draws in short current spikes. On a thin or shared supply the 5 V rail sags and the board resets or hangs mid-render. Pi Zero 2 W: a steady 5 V / 2.5 A supply or better with a short, thick cable; no passive or shared USB hub. Pi 5: the recommended official adapter.
Supply requirements, not a measured draw. No power-draw figure is published for these devices.
A sustained render heats the SoC, and firmware answers by throttling abruptly. A reactive governor caps the clock before that point; --power selects max, balanced (default) or eco. An overclock is opt-in, root-only and off by default.
One engine and no framework on the device. Packages are weightless and sealed; veizik pull fetches the public model and prepares the on-device cache. First activation and periodic re-verification are online; inference itself runs on the board.
| Board | Workload | Measured | Scope |
|---|---|---|---|
| Raspberry Pi 5 | SDXS-512 image | 1.14 GB peak RAM | dep0 native path, warm, whole pipeline |
| Raspberry Pi 5 | Stable Diffusion 1.5, full 25 steps | 1.91 GB peak RAM | warm, whole pipeline |
| Raspberry Pi 5 | SANA-Sprint 0.6B, 1024×1024 | 959 MB peak RAM | warm, 4 threads, whole pipeline |
| Raspberry Pi 5 | Language models, low-memory tier | 104–176 MiB | GPT-2, SmolLM2, LFM2.5 across the tiers shown on the edge page |
| Pi Zero 2 W (512 MB) | GPT-2 124M, sub-byte | 96 MB peak RAM | int8 on the same board: 208 MB |
| Pi Zero 2 W (512 MB) | SDXS-512 prompt to image | ~247 MB peak RAM; 74.9 s | provisional — development build; text encoder, UNet and decoder all on the board |
Figures are veizik’s own on-device measurements on the stated boards. Pack size is a separate axis: SDXS-512 ships as five tiers from 356 MB (quality) to 182 MB (low-memory); Stable Diffusion 1.5 from 967 MB to 496 MB and is Raspberry Pi 5 only, because it does not fit the Zero 2 W.
The parts of the stack that decide whether a job fits and finishes on a small device.
veizik chooses the number format per tensor, so one model ships as five tiers selected with --mode (quality, balanced, fast, ultrafast, low-memory). The formats are proprietary; the measured outcome is shown, not the format internals.
Arm CPU with NEON, Vulkan (V3D), the Hailo-8 NPU through HailoRT, and CUDA on Jetson. There is no separate “mobile” build: the engine executes natively on the board’s cores, built ahead of time for the exact core, with no compile step on the device.
Hailo-8 (26 TOPS) performance results are pending vendor review and are not shown. Not a claim of support for NPUs in general.
A governor samples SoC temperature every two seconds and steps the clock cap down as it warms. Thermal, power and memory are three independent safety axes; each degrades to a supported state instead of crashing.
Power & thermal →Thermal outcomes as measured on the boards, with the missing time-axis and power data stated.
| Device | Measured outcome | Scope |
|---|---|---|
| Pi Zero 2 W | Governor capped at 73.6 °C, held the chip under 76 °C with throttled=0x0; the render completed. An un-governed run goes past 80 °C and resets the board. | One render, SoC temperature as read by the governor |
| Pi 5 (active cooler, stable supply) | 47–78 °C band at the 2.4 GHz native clock, throttled=0x0 in every counted run. | Counted runs; SoC temperature |
| Pi 5 | 16/16 SDXS renders completed, zero resets, no throttling. | On veizik’s board and power supply |
| Boards of one model | Two nominally identical Zero 2 W boards measured about 18% apart. | Thermals and supply dominate; plan per board |
Not measured for this configuration. No time-axis throughput series is published for any edge or mobile device, so this page shows no sustained-throughput figure and no instantaneous peak in its place.
Not measuredNot published. No GPU, SoC, board or wall-power figure, and no joules-per-inference figure with an idle-subtraction basis, exists for these devices. The Jetson AGX Orin is described only as a 60-watt module; that is a module descriptor, not a measured draw.
Not measuredWarm figures are steady-state after the first run. A cold first render and the sealed on-device path are slower and use more peak RAM than the warm figure shown beside each result.
One row per device and workload, each with its own state.
| Class | Device | Runs | State | Evidence scope |
|---|---|---|---|---|
| Board CPU | Raspberry Pi 5 · 4× Cortex-A76 · 8 GB | Language (GPT-2, SmolLM2, LFM2.5); eight computer-vision models | Available now | Measured on the board, CPU only, stock 1.8 GHz unless stated |
| Board CPU | Raspberry Pi 5 | Image (SDXS, SD1.5, SANA-Sprint) and diffusion video (16 frames) | Preview | Render-proven on the board; released one model at a time |
| Board CPU | Raspberry Pi Zero 2 W · 4× Cortex-A53 · 512 MB | Language models (GPT-2, SmolLM2); SDXS-512 image | Language: available now Image: preview | CPU only; the language tier and SDXS-512. The vision set and Stable Diffusion 1.5 need more memory than this 512 MB board. |
| Board CPU + NPU | Raspberry Pi 5 with Hailo-8 (26 TOPS) | Convolution offload through HailoRT, CPU-NEON fallback | Results pending | Performance pending vendor review; not shown |
| Edge GPU | Jetson AGX Orin · 12-core Arm · CUDA · 64 GB | FLUX.1-schnell 12B, 1024×1024, 4-bit weights and 8-bit attention | Preview build, not published | Shown as what the engine draws, not as a reproducible result; no latency published |
| Mobile SoC | Exynos 2600 / Exynos 9825 Galaxy phones | Language and a broad vision set | Deployment pending | Confirmed on the phones; not yet downloadable |
Mobile and Jetson figures that exist (for example GPT-2 124M decode on the Exynos 9825, or MobileNet V1 inference in 4.23 ms) are listed on the edge engineering page as measured with deployment pending; they are not a statement that these devices ship.
LimAxiom and other FPGA or NPU work is in development and not released. It is not part of the matrix above, and nothing here claims support for NPUs in general.
In development / not released--mode tier against measured peak RAM and pack size; a model that does not fit a board (SD1.5 on a Zero 2 W) is not a tuning problem.--power mode (eco for a passive board or marginal supply). Test your own enclosure.Technical demonstrations only; no deployment is implied.