Skip to main content
Checked on Izumi 0.5.7, 2026-11-21: install live on a throwaway stack (Docker Desktop, Mac, arm64), and the requirements and production (Docker, image versions, disk) from the code and by reading. This tab is for the person who installs Izumi on their company’s server: Docker, network, TLS, sign-in and connectors. Commands are allowed. People who only work in Izumi do not need to read it.
Not checked live yet: install on a Linux server (amd64). The throwaway stack ran on a Mac with Docker Desktop (VM with 8 GB). The requirements below come from the code and from the install guide.

What gets installed

Izumi is one docker compose project. The application image is built on your server, there is no registry.
  • app, worker and migrate use the same image team-brain:<version>. The image tag is the version.
  • The image and the repository are still called team-brain, even though the product is called Izumi. Commands run through the script ./tb.

Who runs the install

  • Akia runs the shared and the dedicated install (one instance for several companies, or one instance per client).
  • This tab describes an install run by the company itself, on its own server or in its own cloud. The package and the guide are the same in all three cases.

Hardware

  • Reference server: VPS with 4 vCPU and 8 GB RAM, Linux on amd64.
  • The embedder is the biggest process. docker-compose.yml keeps --max-batch-tokens 512: larger batches get killed for memory right at start on 8 GB.
  • The target is at most 3 GB for the embedder. On a Mac (arm64 image in Docker Desktop) the embedder reached 4.80 GB in an earlier trial, and in the 2026-11-21 test it held 3.6 GiB idle after start, so leave headroom and do not run other large services on the same server.
  • Disk: the team-brain image about 0.5 GB, the database image 0.7 GB, the embedder image about 0.5 GB, the bge-m3 model 2.2 GB in the model volume, and building the image needs about 2 GB more of cache. Plan at least 40 GB for a small install.
  • The database and backups grow with the number of text chunks. 14 daily and 6 monthly copies are kept. The pilot corpus (43 documents, 209 chunks) gives a 1.1 MB backup.
  • Izumi production runs on a server with 75 GB of disk, 26 GB of it used (checked by reading).

Prerequisites

  • Linux on amd64 with Docker Engine and the Docker Compose plugin (docker compose, version 2).
  • The user who runs the install is in the docker group, so docker and ./tb work without sudo. Only reading backup files needs root.
  • openssl (keys and passwords), curl (readiness checks) and git.
  • A DNS name for the server, for example brain.example.com, and open ports 80 and 443 (443 also for UDP) if you use TLS with the proxy profile.
  • Internet during the install: building the image downloads the Node image and npm packages, and the first embedder start downloads the model from huggingface.co. For a server without internet see Running offline.

Limits

  • The reference measurement (embedder RAM and search p95) on a VPS with 8 GB has not been done yet.
  • The install is designed for amd64 (not checked live on a Linux server). ARM (Apple Silicon) is for a trial on a computer, not for a server.
  • Most settings are in the .env file: see Install.

Troubleshooting