Local-first AI runtime for Apple Silicon with CLI and macOS operator workflows for LoRA training, benchmarking, and evaluation.
7
stars
2,930
commits
Swift
primary language
Sep 7, 2026
updated
A native local AI workspace for Apple Silicon — run, fine-tune, and benchmark language models
entirely on your own machine, with no data leaving your desk.
Melix is a complete local AI operations platform built for Apple Silicon Macs. It brings together everything you need to work with language models — serving, fine-tuning, benchmarking, and evaluation — in a single native app and CLI, with no subscriptions, no cloud dependencies, and no data leaving your machine.
Think of it as a local AI studio: a place where you can load a model, adapt it to your needs, measure its quality, and compare results — all from your own hardware.
| Capability | What It Means For You |
|---|---|
| Model Registry | Import models from disk or download from Hugging Face into a local registry you control |
| Server Sessions | Start, pause, resume, and stop local model servers with a single command or click |
| Local Chat | Chat with any registered model from the menubar app or CLI, staying fully offline |
| LoRA Fine-Tuning | Train a custom adapter on your own dataset to specialize a model for your use case |
| Benchmarking | Run repeatable quality benchmarks and matrix comparisons across models and adapters |
| Evaluation | Score models against standard suites (MMLU and more) and export the results |
| Native macOS App | A polished menubar and workspace UI that puts all of the above one click away |
Local AI work tends to scatter across too many tools: one script for serving, another for training, a notebook for evaluation, and a shell history for benchmarks. Melix keeps that entire loop in one place, on one machine.
Make sure you have macOS 15+ on Apple Silicon, plus swift, python3, and uv installed.
# 1. Bootstrap the repository
make bootstrap
# 2. Generate protocol artifacts
make proto
# 3. Run verification gates
make swift-test
make py-test
make integration-test
Then bring up the local stack:
bash scripts/dev_up.sh
For the full native macOS experience:
bash scripts/dev_app_up.sh
New here? The Getting Started guide walks through every step with context.
| Document | What's Inside |
|---|---|
| Getting Started | Step-by-step setup from a fresh checkout to a working loop |
| Current Status | What is shipped today and where the honest boundaries are |
| Phase Roadmap | The original phase model and its current completion state |
| Contributing | How to contribute, what to verify, and what a good PR looks like |
| Docs Map | Full index of specs, runbooks, architecture decisions, and plans |
| Marketing & Storytelling Kit | Product overview, LoRA narrative, reusable copy, and screenshot evidence |
| Benchmark & LoRA Runbook | Deep-dive operator guide for benchmarking and fine-tuning |
| Local Install Runbook | Install Melix as a persistent local service |
| Packaging Targets | Homebrew, launch agent, and app-bundle delivery options |
Contributions are welcome — documentation, tooling, protocol, runtime, CLI, and macOS operator surface. Start with docs/contributing.md for the workflow, verification commands, and handoff expectations.
The short version:
main.make swift-test && make py-test && make integration-test before opening a PR.Melix is licensed under the Apache License, Version 2.0. See LICENSE.
Swift
54.0%
Python
45.8%
Local-first AI runtime for Apple Silicon with CLI and macOS operator workflows for LoRA training, benchmarking, and evaluation.
7
stars
2,930
commits
Swift
primary language
Sep 7, 2026
updated
A native local AI workspace for Apple Silicon — run, fine-tune, and benchmark language models
entirely on your own machine, with no data leaving your desk.
Melix is a complete local AI operations platform built for Apple Silicon Macs. It brings together everything you need to work with language models — serving, fine-tuning, benchmarking, and evaluation — in a single native app and CLI, with no subscriptions, no cloud dependencies, and no data leaving your machine.
Think of it as a local AI studio: a place where you can load a model, adapt it to your needs, measure its quality, and compare results — all from your own hardware.
| Capability | What It Means For You |
|---|---|
| Model Registry | Import models from disk or download from Hugging Face into a local registry you control |
| Server Sessions | Start, pause, resume, and stop local model servers with a single command or click |
| Local Chat | Chat with any registered model from the menubar app or CLI, staying fully offline |
| LoRA Fine-Tuning | Train a custom adapter on your own dataset to specialize a model for your use case |
| Benchmarking | Run repeatable quality benchmarks and matrix comparisons across models and adapters |
| Evaluation | Score models against standard suites (MMLU and more) and export the results |
| Native macOS App | A polished menubar and workspace UI that puts all of the above one click away |
Local AI work tends to scatter across too many tools: one script for serving, another for training, a notebook for evaluation, and a shell history for benchmarks. Melix keeps that entire loop in one place, on one machine.
Make sure you have macOS 15+ on Apple Silicon, plus swift, python3, and uv installed.
# 1. Bootstrap the repository
make bootstrap
# 2. Generate protocol artifacts
make proto
# 3. Run verification gates
make swift-test
make py-test
make integration-test
Then bring up the local stack:
bash scripts/dev_up.sh
For the full native macOS experience:
bash scripts/dev_app_up.sh
New here? The Getting Started guide walks through every step with context.
| Document | What's Inside |
|---|---|
| Getting Started | Step-by-step setup from a fresh checkout to a working loop |
| Current Status | What is shipped today and where the honest boundaries are |
| Phase Roadmap | The original phase model and its current completion state |
| Contributing | How to contribute, what to verify, and what a good PR looks like |
| Docs Map | Full index of specs, runbooks, architecture decisions, and plans |
| Marketing & Storytelling Kit | Product overview, LoRA narrative, reusable copy, and screenshot evidence |
| Benchmark & LoRA Runbook | Deep-dive operator guide for benchmarking and fine-tuning |
| Local Install Runbook | Install Melix as a persistent local service |
| Packaging Targets | Homebrew, launch agent, and app-bundle delivery options |
Contributions are welcome — documentation, tooling, protocol, runtime, CLI, and macOS operator surface. Start with docs/contributing.md for the workflow, verification commands, and handoff expectations.
The short version:
main.make swift-test && make py-test && make integration-test before opening a PR.Melix is licensed under the Apache License, Version 2.0. See LICENSE.
Swift
54.0%
Python
45.8%