Brethof AI Reviews

LocalAI (self-hosted, Docker) review

Your own AI server, with a model gallery built in

Not sponsoredModels & AI Infrastructure

Tested on Linux · Free · Open source

LocalAI is a free server that runs open models on your hardware, with OpenAI, Anthropic and Ollama APIs. We ran its CUDA 12 Docker image on Linux, with Qwen3.5 9B on an RTX 4060 Ti.

The good

  • Our model, in the gallery. Our exact model is in its gallery, but the main Install button picks a faster variant. Ours is the Base build row, a click more.
  • Whole model on GPU. With the CUDA image, the whole model went onto the GPU, 8,192 token context, no settings touched. Its first speed reading counts the load.
  • No account, MIT. It opens straight to its home page, no login. No telemetry was found in its code. It is MIT licensed, no paid plans.

The not so good

  • Big, built for operators. Two big downloads come first, a 3.74 gigabyte image, then a 1.92 gigabyte backend from a moving tag, unsigned, silent at 100 percent for two minutes. Fifty pages is a lot.
  • Whole PDF, no citations. The chat pastes a PDF's whole text into your message, without retrieval or citations, and refuses scanned PDFs. It is weeks old. Real retrieval needs Agents and an embedding model.

Our verdict

Recommended with caveats

MIT, no account, no telemetry, and our exact model from its gallery ran fully on the GPU. But it listens on every network with no key, and silently installs an unsigned backend from a moving tag. Built for operators.

Reviewed: LocalAI (self-hosted, Docker) v4.11.0 (CUDA 12 image) · facts checked 2026-10-22

Watch it onYouTube

Not sure which tool fits your setup? Tech advice: ask us before you spend yours. Describe your situation and get our answer in writing within two working days.

Ask us →

← All reviews

Hear it when it ships

New releases, real benchmarks and the occasional deep-dive. No spam, unsubscribe in one click.

Everything we build