Documentation Coming soon

Dunesies Personal
From download to first answer.

Hi
Hello, I'm Dunesies. I run locally on your device.
Nice, so nothing I say leaves my Mac?
Exactly. No cloud, no accounts. It all stays right here.
How much does this cost?
Dunesies itself is free. You just need your own Mac to run it on.
Can I switch models anytime?
Anytime. Download any model you like and switch in seconds.
Send a message

Requirements

What you need to run it.

macOS macOS 14.0 or laterUniversal build for Apple Silicon and Intel.
iOS iOS 17.0 or lateriPhone and iPad, with smaller models sized to fit.
Memory 8 GB minimum, 16 GB recommendedmacOS reserves 2.5 GB on Apple Silicon and 3.5 GB on Intel when rating a model.
Disk space About 2 GB, plus each modelModels range from 1.6 GB to 40 GB.
Internet Setup and downloads onlyInference itself is fully offline.
Cost FreeNo account, no subscription, no per token billing.

Getting started

Eight steps, and only the first two happen more than once.

01

Install the app

Open the .dmg and drag Dunesies into your Applications folder.

02

Set up the engine

This is automatic. The app downloads a prebuilt llama.cpp engine on first launch, around thirty seconds, once.

03

Pick a model that fits your machine

The Models tab rates every model against your real hardware, so start with anything marked Runs Great. 7B or 8B is the sweet spot on 16 GB machines.

04

Download it once

Click download and wait for the file to land, paste a Hugging Face GGUF link, or choose a .gguf file already on your disk.

05

Start a chat

Switch to the Chat tab, choose your model from the pill at the top, and type. Answers stream back as they generate. The plus button starts a new session.

06

Bring in your own files

Use the paperclip to attach a PDF, source file, CSV or log and ask about its contents. Vision models accept images too.

07

Work offline

Once a model is downloaded you can disconnect entirely. On a plane, on a train, or on a locked down network, it works exactly the same.

08

Switch models whenever you like

Keep several models installed and change between them from the picker at the top of the chat. Each session remembers what it was started with.

No terminal required. There is no Python to install, no Xcode tools and no compiler prompts. The engine is fetched prebuilt on first launch.

Models

A model is a single .gguf file on your disk. You download it once, and it stays exactly as it is: no silent updates, no retirement notice, no change in answers overnight.

The Models tab has four sources:

Model Library

Every model you have already installed. Pick one and start a chat.

Suggested

Nineteen ready to run models, one click each. This is the tab to use first.

Custom URL

A direct Hugging Face link to a .gguf file, for anything outside the suggested list.

Local File

A .gguf file already sitting on your disk. Nothing is downloaded.

Downloading from Hugging Face

Open the Custom URL tab and paste a direct link to the .gguf file. It has to be the direct file link, the one containing resolve/main, not the model’s landing page.

https://huggingface.co/username/model/resolve/main/file.gguf
01

Find the GGUF file on Hugging Face

Open the model repository, go to Files and versions, and choose the quantisation you want. Q4_K_M is what every suggested model uses.

02

Copy the download link

Use the file’s download link rather than the page URL. It contains resolve/main and ends in .gguf.

03

Paste it into Custom URL

The model name is optional. Leave it empty and it is taken from the URL.

04

Download once

The file lands on your disk and is yours. Everything after this works with no connection.

It must be a GGUF file. Repositories publishing only safetensors or PyTorch weights will not load. Look for a repository with GGUF in its name, or a GGUF conversion of the model you want.

Choosing a model for your device

You do not have to work this out yourself. The app reads your actual memory and chip and labels every model before you spend time downloading it. This is how that label is decided.

First it sets aside memory for the system, then rates the model against what is left:

Apple Silicon Reserves 2.5 GB, runs on GPU and CPURuns Great up to 60% of the remainder, May Be Slow up to 95%, Not Recommended above that
Intel Reserves 3.5 GB, runs on CPU onlyRuns Great up to 50% of the remainder, May Be Slow up to 80%, Not Recommended above that

What that means in practice

Your machineUsableRuns Great up toA good first model
8 GB Apple Silicon5.5 GB3.3 GBGemma 2 2B or Qwen2.5 3B
16 GB Apple Silicon13.5 GB8.1 GBLlama 3.1 8B or Qwen2.5 7B
32 GB Apple Silicon29.5 GB17.7 GBGemma 2 27B or Qwen2.5 14B
64 GB Apple Silicon61.5 GB36.9 GBQwen2.5 32B or Mixtral 8x7B
16 GB Intel12.5 GB6.3 GBQwen2.5 7B or Mistral 7B
A bigger model is not automatically a better answer. A 7B or 8B model is the sweet spot on a 16 GB machine: it answers quickly and handles most work. Reach for a larger one when you have a task the smaller model is genuinely getting wrong, not by default.

The suggested models

Nineteen models, one click each. Every one is Q4_K_M except Phi-3 Mini, so sizes are directly comparable.

ModelParametersDownloadQuantBest for
Gemma 2 2B Instruct2B1.6 GBQ4_K_MQuick responses, low resources
Qwen2.5 3B Instruct3B1.9 GBQ4_K_MGeneral chat, quick responses
Llama 3.2 3B Instruct3B2.0 GBQ4_K_MGeneral purpose, fast
Phi-3 Mini 3.8B3.8B2.2 GBQ4Reasoning, coding, chat
Phi-3.5 Mini 3.8B3.8B2.3 GBQ4_K_MReasoning, efficient performance
Qwen2.5 7B Instruct7B4.4 GBQ4_K_MGeneral purpose, versatile
Mistral 7B Instruct v0.37B4.4 GBQ4_K_MChat, reasoning, coding
Qwen2.5 Coder 7B7B4.7 GBQ4_K_MCoding, lightweight
Llama 3.1 8B Instruct8B4.9 GBQ4_K_MGeneral purpose, versatile
Gemma 2 9B Instruct9B5.4 GBQ4_K_MReasoning, math, science
CodeLlama 13B Instruct13B7.9 GBQ4_K_MCode generation, debugging
Qwen2.5 14B Instruct14B8.5 GBQ4_K_MAdvanced chat, analysis, coding
Gemma 2 27B Instruct27B16 GBQ4_K_MAdvanced tasks, high quality
Qwen2.5 32B Instruct32B19 GBQ4_K_MProfessional tasks, complex reasoning
Qwen2.5 Coder 32B32B19 GBQ4_K_MAdvanced coding, debugging, refactoring
DeepSeek Coder 33B33B19 GBQ4_K_MProfessional coding, architecture design
Mixtral 8x7B Instruct47B (8x7B MoE)26 GBQ4_K_MEfficient large model, strong general reasoning
Nous Hermes 2 Mixtral 8x7B47B (8x7B MoE)26 GBQ4_K_MEnhanced chat, instruction following
Llama 3.3 70B Instruct70B40 GBQ4_K_MProfessional use, complex reasoning

Working with your files

The documents you would never upload anywhere.

Attach what you are working on

PDFs, source files, CSVs, logs and markdown. Images too, when a vision model is loaded.

Knowledge that stays with a session

A session can carry its own set of documents, so a work session keeps your project docs and a study session keeps your textbook.

Isolated sessions

Every conversation runs in its own process with its own history, model and settings. Context never bleeds between chats.

Searchable history

Local search across every session you have ever had, with one click switching between conversations.

Privacy and offline use

Proof, not a promise.

Fully local inference

Models run on your machine through llama.cpp. Nothing is uploaded, nothing is logged remotely.

No account, no telemetry

There is no sign up, no subscription to cancel and no telemetry to opt out of.

Verify it yourself

Disconnect the network mid conversation and keep going. That is the only proof of privacy that means anything.

Offline after download

Once a model is on disk, no connection is needed at all.

Support

Something not covered here, or not behaving as described? Tell us what you were doing and what happened.

Contact us Try Dunesies