Architecture
What it takes to run Senua AI.
Senua AI runs locally: on premise, on the consumer hardware you already own, and all the way down to edge devices and wearable-class silicon. No cloud dependency, no GPU, no model too big to carry. This page is the proof: what it compiles to, what it needs, and what it costs to run: measured on the current build, checkable against our datasheets and validation notes. We don’t publish what’s inside (that’s patent-pending and stays ours) : and if these numbers look impossible for a neural network, that’s the point: it isn’t one.
Measured footprint
The numbers, as built.
Taken from the current development build: not projections.
9.3 MB
The entire cognition engine (reasoning, learning, memory, language) compiled to ONE native binary. The web console adds 0.5 MB; the CLI 0.2 MB. A full install is under 15 MB.
One binary, no runtime
Disk = the learned store
An Anima’s footprint matches what it has learned: there is no fixed model file. As a guide only: the current 1.6M-fact reference Anima is 2.2 GB on disk. Footprint scales with the corpus you feed it, and nothing else.
Guide: 1.6M facts ≈ 2.2 GB
4 vCPU
The cloud reference (cloud-1): ONE general-purpose node: 4 vCPU, 16 GB RAM, 500 GB disk : runs the full platform and is staging a 117 GB raw public corpus (866M facts) toward a large-scale Anima.
m7i.xlarge class · guide only
0 GPUs
No GPU, no accelerator, and no floating-point unit required: the arithmetic is integer-only end to end. No linear algebra, no per-token cloud bill.
CPU-native
2 vCPU arm64
The autonomous-flight reference: full missions flown from a 2-vCPU arm64 node (AWS Graviton). The onboard target is a quad-core Cortex-A53 with 512 MB RAM (Pi Zero 2 W class) inside a 1.3 W fly-and-monitor battery envelope: the one running metric that matters at the edge.
Battery envelope: see DS-001
2 architectures
Native machine code for x86_64 and arm64: the arm64 port is hardware-validated on AWS Graviton. macOS, Linux, and bare metal via SenuaOS.
x86_64 + arm64
Baseline hardware
Yours, not ours: the hardware ladder.
The floor is lower than you think, and every rung is hardware you own. Wearable-class silicon: a quad-core arm64 chip with 512 MB of RAM (Raspberry Pi Zero 2 W class: the same silicon class as today’s wearables) runs a device leaf, flying and monitoring a drone inside a 1.3 W power envelope. Consumer hardware: an ordinary laptop or a Pi CM5-class module runs the full assistant, console and all: private, offline, no account required. On premise: one commodity server puts the whole platform inside your firewall, air-gapped if you need it. Cloud, if you choose it: a single 4-vCPU general-purpose node serves the complete platform and builds the Anima. No cluster, no accelerator fleet, no special hardware anywhere in the stack: local is not the compromise; it is the design.
Proving grounds
Tested where you can check it.
The platform is exercised on AWS test harnesses that mirror real deployments: an arm64 Graviton flight rig where Senua AI flies full autonomous missions against a real autopilot with QGroundControl recording the ground truth; power and battery simulations cross-implemented in three independent code stacks with byte-identical outputs; and live host-metrics monitors across macOS, x86_64 Linux, and arm64 Linux. The results are published below as they land.
Datasheets
Single-page engineering references: power architectures, deployment envelopes, hardware baselines. PDF and SVG, with their assumptions stated.
DS-series
Validation notes
How each published number was produced and independently cross-checked, so a skeptical engineer can rerun it rather than trust it.
VN-series
Latest releases
New datasheets, validation notes, and platform milestones, newest first, in the News section.
News
Want the envelope for your hardware?
Tell us the board, the box, or the instance type: we’ll tell you what fits. The business case is over at Why Senua.
Request access