features

Everything runs here.

No part of emb3r depends on a service staying up, a key staying valid, or a company staying in business.

the short version

Twelve things that fall out of running the model on your own hardware — plus a couple you get to switch on yourself.

Fully offline

Works on a plane, in a SCIF, behind a firewall. After the first model download, emb3r never opens a socket.

Watch it think

Replies stream in as they're generated, with a stop button if it's rambling and a live tokens/sec and context readout — not a blank wait, then a wall of text.

Fast on modest hardware

Quantized runtime with a Metal backend on Mac and Vulkan or CPU on Windows. 4 GB of RAM is enough to get going.

Reads your files

Attach a text file and ask about it. The contents are read on your machine and never leave it.

Matched to your machine

On first launch emb3r reads your RAM and CPU and recommends a model that will actually run — no guessing, no failed loads.

Remembers your chats

Each profile keeps its own conversation history, saved to disk and restored on launch — with real memory of what was said, not just old text on screen.

A shelf of models

Choose from a curated set of open models — Llama, Qwen, Mistral — and switch between them in Settings.

Shape its personality

The instructions that make Ember, Ember are editable in Settings — not hardcoded. Make it formal, make it terse, make it something else entirely.

Open on GitHub

The full source is public. Read it, fork it, audit exactly what runs. No accounts, no seats, no phone-home.

Web access — off unless you turn it on

Bring your own free Gemini key and emb3r can look things up for questions that need current information. Only that one question goes out; everything else stays local. If Gemini's ever unavailable — rate-limited, whatever — emb3r just answers locally instead of leaving you stuck. The model itself is overridable in Settings too, for accounts that need a specific one. Flip it off in Settings and it's back to fully offline.

Updates without leaving

emb3r checks for new releases on launch and lets you download and install them from Settings — no browser, no manual reinstall.

Make it look like yours

Pick your own accent color and light or dark theme. emb3r keeps whatever you choose readable — it won't let you pick a color you can't actually read.

the models

Six open models, all 4-bit quantized (Q4_K_M) GGUF weights pulled from Hugging Face. emb3r starts you on the smallest — you can download any of the others from Settings whenever you want.

Llama 3.2 3B Instruct · default
2.0 GB · needs 4 GB RAM
Qwen2.5 3B Instruct
1.9 GB · needs 4 GB RAM
Mistral 7B Instruct v0.3
4.4 GB · needs 8 GB RAM
Qwen2.5 7B Instruct
4.7 GB · needs 8 GB RAM
Llama 3.1 8B Instruct
4.9 GB · needs 8 GB RAM
Qwen2.5 14B Instruct
9.0 GB · needs 16 GB RAM

emb3r won't let you download a model your machine can't hold — if you have 8 GB of RAM, the 14B simply isn't offered. Bigger models are slower to answer but better at longer reasoning.

system requirements

Minimums for the smallest model. Larger models want more RAM and disk.

Memory
4 GB RAM minimum (8 GB+ for bigger models)
Disk
1.9–9 GB per model
macOS
Apple silicon & Intel
Windows
10 / 11 · x64
Acceleration
Metal on Mac · Vulkan or CPU on Windows