Requirements and sizing

The Aptabit Platform runs on a Linux server owned by your organization, physical or virtual. The values below come from the CLI's own checks. Production sizing depends on the number of people, the volume of documents, and the local models you choose, and is defined per project with Aptabit.

Operating system

Item	Requirement
System	Linux on a server or virtual machine.
Architecture	x86_64 or arm64. Other architectures are blocked by the machine check.
Distributions	The CLI is a static binary, with no glibc or OpenSSL dependency. It runs on any x86_64 or arm64 Linux, including RHEL 8, Ubuntu 20.04, and Debian 11.
Other systems	macOS for development only. Windows is not supported, and WSL triggers a warning.

Memory and disk

Resource	Minimum	Recommended
RAM	4 GiB	8 GiB
Free disk	10 GiB	50 GiB

Below the minimum, aptabit check blocks the installation. Without local models, the platform takes about 4.6 GB of images and 0.7 GiB of RAM. Each local model adds its own memory needs (RAM or video memory), and aptabit check reports the largest catalog model that fits the machine.

Container runtime

Docker Engine with Compose v2: the default path of the guided installer. docker-compose v1 on its own blocks the installation.
Podman: partial support, through advanced configuration.
The user running the CLI needs access to Docker. The check also detects Docker installed as a snap, rootless mode, and SELinux.

Accelerators

Hardware	How it is used
CPU	The platform works without an accelerator. llama.cpp is the default runtime for local models; on Intel processors, OpenVINO also runs on the CPU.
NVIDIA GPU	Through CUDA, with the NVIDIA driver and the NVIDIA Container Toolkit installed. Without the toolkit, the platform starts on the CPU and the check names what is missing.
Intel GPU	Integrated or discrete, through `/dev/dri`, in the OpenVINO and IPEX-LLM runtimes.
AMD GPU	Detected by the check but not used yet: models run on the CPU.
NPU	Model execution on NPUs is available.

aptabit check lists every GPU and NPU on the machine, with memory, driver, and Docker access. The vLLM, TensorRT-LLM, and SGLang runtimes require a CUDA GPU. See Models, runtimes, and accelerators.

Dependencies

Each dependency can run with the platform or point to a service your organization already operates:

Dependency	In the stack	External
Relational database	PostgreSQL 16	Managed PostgreSQL
Sessions and queue signals	Redis 7	Managed Redis
Object storage	MinIO	S3-compatible service
Vector database	Milvus 2.4 with etcd	Milvus or Zilliz

See Operations.

Network

Port: the platform's proxy serves HTTP on one port (8080 by default).
HTTPS: terminate TLS at your organization's load balancer or reverse proxy.
Connected installation: outbound HTTPS to Aptabit Manager, to Aptabit's private registry, and to the public registries of the base images (Docker Hub and quay.io), with the clock within one minute of Aptabit Manager's clock. Managed local models also pull runtime images and weights from Hugging Face. See Other outbound connections.
Isolated installation: no outbound traffic to Aptabit. See License and connectivity.
Corporate proxy: the CLI trusts the system's certificate authorities. A proxy that inspects HTTPS works once its CA is in the system trust store.

Interface languages

The web interface is available in nine languages, with Portuguese, English, and Spanish as the most complete translations, and offers light and dark themes. The CLI speaks Brazilian Portuguese and English.