Hardware Compatibility¶

Effero is designed to run on hardware ranging from a Raspberry Pi to a rack-mounted server. This table tracks what has been tested.

Help wanted: if you run Effero on hardware not listed here, please open a PR adding your measurements.

Tested Platforms¶

Device	CPU	RAM	OS	Effero Version	Status	Cold Start	Idle RAM	Notes
x86-64 desktop	i7-12700 / Ryzen 7	32 GB	Windows 11 / Ubuntu 24.04	0.1.9+	✅ Tested	~1.6s (warm) / ~4.5s (cold)	45.9 MB	Primary development target; measured on Python 3.13
Raspberry Pi 5	Cortex-A76	8 GB	Bookworm	—	🔲 Untested	—	—	Target platform, needs validation
Jetson Orin Nano	Cortex-A78AE	8 GB	JetPack 6	—	🔲 Untested	—	—	GPU acceleration for vision/VLA
Generic x86 mini-PC	N100	16 GB	Ubuntu 24.04	—	🔲 Untested	—	—	Budget always-on home server

What we measure¶

Cold Start: Time from effero run to first skill invocation ready (no model loading).
Idle RAM: Resident set size with the agent running, no active skills, no model loaded.
Model loading: Separate — depends entirely on the model backend (Ollama, llama.cpp, cloud).

Minimum requirements (estimated)¶

Python 3.11+
512 MB RAM (core runtime only, no models)
ARM64 or x86-64
Linux, macOS, or Windows (WSL2)

On-Device Inference Engine (Cactus) Footprint¶

When using Effero with on-device LLM inference via the Cactus runtime (backend: cactus), quantized models run locally with sub-100ms first-token latencies:

Model Architecture	Quantization	RAM Footprint	Device Target	Latency (First Token)
1B-3B Edge SLM	INT4	~1.2 GB - 2.1 GB	Raspberry Pi 5 / Mobile / Jetson Nano	< 60 ms
7B-8B Compact LM	INT4	~4.2 GB - 4.8 GB	Jetson Orin Nano / N100 Mini-PC	< 120 ms
7B-8B Compact LM	INT8	~7.8 GB - 8.5 GB	x86 Desktop / Jetson AGX	< 90 ms

Confidence-Based Hybrid Cloud Handoff¶

Effero supports hybrid routing with Cactus: 1. The on-device SLM produces local reasoning and an action plan with an evaluated confidence score. 2. If confidence meets or exceeds model.min_confidence (default 0.70), the action is dispatched locally with zero external API latency or cost. 3. If the model encounters ambiguity or confidence falls below the threshold, Effero automatically escalates the context to the configured cloud fallback backend (e.g. openai:gpt-4o, anthropic:claude-sonnet-4-5).

Google AI Edge LiteRT-LM (In-Process NPU & Edge Engine)¶

For single-board computers (Raspberry Pi 5 with AI HAT), robotics controllers, and edge Linux boards, Effero supports Google LiteRT-LM (backend: litert).

Unlike background HTTP server daemons, LiteRT-LM runs in-process via native Python bindings, directly accessing device NPUs and GPUs:

Hardware Target	Accelerator Backend	Model Format	Recommended SLM
Raspberry Pi 5 + AI HAT	NPU / Hailo-8 / Coral	`.litertlm`	Gemma 3 1B / SmolLM 1.7B
Jetson Orin Nano / AGX	GPU (Vulkan/OpenCL)	`.litertlm`	Gemma 3 4B / Qwen 2.5 7B
Android Robotics Tablets	NNAPI / Hexagon NPU	`.litertlm`	Gemma 3 1B / MobileLM
Standard Edge x86/ARM	Multi-threaded CPU	`.litertlm`	Compact SLMs