---
title: README
emoji: 🔒
colorFrom: gray
colorTo: green
sdk: static
pinned: false
---

<div align="center">

# HilbertRaum

### Your private AI workspace, fully offline

Chat with a local AI, ask questions about your own documents, and keep everything on your computer.
No cloud, no account, no telemetry.

[**Download**](https://github.com/HilbertraumAI/HilbertRaum/releases/latest) ·
[**Source on GitHub**](https://github.com/HilbertraumAI/HilbertRaum) ·
[**Website**](https://hilbertraum.ai) ·
[**Verified model catalog**](https://huggingface.co/collections/HilbertRaum/hilbertraum-verified-model-catalog-6a9f0628520b64acd0e93fea)

</div>

---

## What HilbertRaum is

HilbertRaum is an open-source desktop app (Windows, macOS, Linux; GPL-3.0-or-later) that runs open-weight
GGUF models with `llama.cpp` entirely on your machine. It ships document Q&A with citations, evidence
review, image understanding, Whisper transcription and dictation, offline translation for 51 languages,
and an optional OpenAI-compatible loopback API for other programs on the same computer.

The whole workspace is portable and encrypted: app, models and data can live on a USB stick or an
external SSD and move between machines.

## How we pick models

The catalog is deliberately small. Every model is tested offline on real laptops for answer quality
(a grounded question-answering set in German and English, scored without a cloud judge), decode speed
and peak RAM, and the app recommends one model for your hardware so you never have to compare quants
yourself. Weights are verified by SHA-256 before use.

| Your RAM | Recommended chat model |
| --- | --- |
| 8 GB | Qwen3.5 4B (UD-Q4_K_XL) |
| 12 to 15 GB | Gemma 4 E2B Instruct QAT Q4_0 |
| 16 to 23 GB | Qwen3.5 9B (UD-Q4_K_XL) |
| 24 GB | Qwen3.8 27B UD-Q4_K_M |
| 32 GB and up | Qwen3.8 27B UD-Q5_K_M |

Supporting models: Multilingual E5 Small (embeddings), BGE Reranker v2 M3, Whisper Small,
Qwen2.5-VL 3B (vision), TranslateGemma 12B (translation, opt-in).

The full list with sizes, minimum RAM and licenses is in the
[verified model catalog](https://huggingface.co/collections/HilbertRaum/hilbertraum-verified-model-catalog-6a9f0628520b64acd0e93fea)
and in the [README on GitHub](https://github.com/HilbertraumAI/HilbertRaum#supported-models).
The benchmark protocol and measured numbers are in
[`docs/model-benchmarks.md`](https://github.com/HilbertraumAI/HilbertRaum/blob/main/docs/model-benchmarks.md).

## What is coming here

- The grounded-QA evaluation set and benchmark results as datasets
- SHA-pinned copies of the catalog weights with our measurements in each model card

Questions and bug reports go to [GitHub Issues](https://github.com/HilbertraumAI/HilbertRaum/issues).