Can my Mac actually run an LLM?

Yes. Any Mac with Apple Silicon (M1 or later) can run a local LLM. TARX uses Metal GPU acceleration to run inference directly on your Mac's unified memory. An M1 MacBook Air with 16GB RAM handles it fine. More RAM and newer chips mean faster responses.

How much storage does a local LLM need?

TARX's default model is about 4.7 GB (quantized to Q4_K_M). That's comparable to a large app. You need roughly 10-15 GB free for comfortable operation including the app, model, and local embeddings.

Is local inference slower than cloud AI?

For the same model size, local inference on Apple Silicon is competitive with cloud. You lose network latency (which can be 200-500ms per request to cloud APIs). For most conversational and coding tasks, TARX responds in under 2 seconds on M1 and under 1 second on M3/M4.

How does TARX compare to Ollama or LM Studio?

Ollama and LM Studio are model runners — they give you a CLI or UI to run open models. TARX is a full AI platform: it runs inference, fine-tunes on your usage, indexes your files, connects to a peer-to-peer Supercomputer mesh, and integrates with Xcode and VS Code. Think of Ollama as the engine; TARX is the car.

Can I use my own models with TARX?

TARX manages its own model lifecycle, but the underlying format is GGUF (llama.cpp compatible). Advanced users can swap models via the Bridge configuration. The Supercomputer mesh supports multiple model architectures.

Does running a local LLM drain my battery?

During active inference, yes — GPU compute uses power. But TARX only runs inference when you're interacting with it. Idle power draw is near zero. On a MacBook Pro M3, expect roughly the same battery impact as running Xcode builds.

Run a Local LLM on Your Mac — No Cloud, No Compromise

guide5 min read·by TARX

Your Mac is a modeling machine. Apple designed it that way — unified memory, Metal GPU, Neural Engine. Everything a language model needs to run, train, and improve is sitting on your desk right now.

And you're sending your questions to OpenAI's data center in Iowa. Let's fix that.

TARX turns your Mac into a personal modeling layer — it runs AI inference locally, fine-tunes on your usage patterns, and gets better at your work every day. No cloud. No API keys. No one watching.

The state of local AI on Mac in 2026

Apple Silicon changed what's possible. The M1, M2, M3, and M4 chips have unified memory architectures — the CPU and GPU share the same RAM. This means a 32GB MacBook Pro can load a model that would require a dedicated GPU card on any other platform.

The ecosystem has caught up:

llama.cpp compiles natively for Metal, achieving near-optimal token throughput
MLX (Apple's ML framework) enables native training and inference on Apple Silicon
GGUF format has become the standard for quantized model distribution
FoundationModels.framework (iOS 26+) provides system-level on-device inference

You don't need a Linux box. You don't need an NVIDIA GPU. You don't need a cloud account. Your Mac is a capable inference machine right now.

What TARX does with that hardware

TARX is built for this reality. It's not a chatbot wrapper. It's a modeling layer that manages the entire lifecycle — from first inference to continuous fine-tuning:

Model management. TARX downloads, quantizes, and serves models via llama-server with Metal acceleration. The default model runs at Q4_K_M quantization — small enough to fit in memory alongside your other apps, accurate enough for professional work.

Inference. When you send a message, TARX runs inference on your Mac's GPU. No network request. No API key. No rate limit. Response time depends on your hardware — M4 Pro is fastest, M1 is still perfectly usable.

The modeling layer. This is where TARX diverges from every other local AI tool. Ollama runs a model. LM Studio runs a model. TARX models you. It continuously captures your usage patterns and trains LoRA adapters via MLX. A developer's TARX learns their codebase conventions. A writer's TARX learns their voice. A compliance officer's TARX learns their regulatory framework. The fine-tuning happens locally, on your hardware, automatically. Your model gets better at your work every day — and you can see how.

File indexing. Drop files into a TARX Space and they're indexed locally using embeddings (nomic-embed-text). TARX retrieves relevant context per query — like RAG, but everything stays on your device.

The Supercomputer. When you need more than your local hardware can provide, TARX connects to a peer-to-peer mesh of Mac hardware. Not a cloud. A mesh. Your query routes to other Macs running TARX. You can contribute your hardware and earn credits.

Setup — under 5 minutes

Download: Go to tarx.com/download. Grab the DMG. Drag to Applications.
Launch: Open TARX. The model downloads automatically (~4.7 GB, one time). No account creation. No API key entry. No email verification.
Use: Start typing. TARX is running local inference on your GPU. The green dot in the UI means you're on the Supercomputer. No dot means local-only. Both work.
Add context: Drag project folders or files into a Space. TARX indexes them and uses the context in every response.

That's it. You're running a local LLM on your Mac.

Hardware requirements

| Mac | RAM | Experience | |-----|-----|------------| | M1 MacBook Air, 16GB | 16GB | Good. ~15 tok/s. Comfortable for conversation and light coding. | | M1 Pro/Max, 32GB | 32GB | Great. ~25 tok/s. Room for larger contexts alongside other apps. | | M2/M3/M4 MacBook Pro, 32-64GB | 32-64GB | Excellent. 30-50+ tok/s. Handles complex multi-file reasoning. | | M4 Max Mac Studio, 128GB | 128GB | Extreme. Could run 70B models locally. |

Minimum: M1 with 16GB. Recommended: M2+ with 32GB. TARX runs on any Apple Silicon Mac.

The privacy question

Every cloud AI service has a terms-of-service clause about your data. Some promise not to train on it. Some don't. All of them can see it — your prompts, your code, your documents pass through their infrastructure.

With TARX, the question doesn't exist. The model is a file on your SSD. Inference is a GPU computation on your hardware. There is no server to log your queries. There is no database storing your prompts. There is no third party.

This isn't a privacy policy. It's physics. Your data can't leak to a server that doesn't exist.

Beyond the basics

Once you're running local, the possibilities open up:

Spaces — organize your AI context by project, client, or topic
Skills — automate workflows (Xcode builds, code review, research pipelines)
Developer API — OpenAI-compatible endpoints on localhost for your own tools
MCP Server — connect TARX to Claude Code, VS Code, and other MCP-capable tools
Supercomputer — contribute your hardware to the mesh and earn credits

Your Mac is more capable than you think. TARX is how you use it.

How TARX models your work — transparently

Other local AI tools are model runners. They give you a command line, you pick a GGUF, you chat. The model is the same on day 100 as it was on day 1.

TARX is a modeling layer. It captures your patterns, trains adapters, deploys improvements, and — soon — shows you what it learned. The modeling receipt: a transparent account of how your usage shaped the model. Which patterns it absorbed. Where its confidence grew. How the model changed from the generic weights you downloaded to the specialized tool it became.

This isn't a feature you toggle. It's the architecture running underneath every conversation. When the transparency surface ships, you'll see your model's trajectory. Until then, the flywheel is already turning — every conversation trains, every adapter deploys, your TARX gets better at you.

This page has a sector model

This page is connected to TARX's developer sector model. When the embedded TARX conversation activates here, you'll be talking to a model fine-tuned on developer-specific patterns: Apple Silicon inference, local model management, build systems, performance tuning.

Every conversation on this page feeds the developer model. The model serves the page. The page trains the model. That's the flywheel.

Five sectors are in development: developer, enterprise, compliance, gov, and SMB. This page is the developer surface.