ToshLLM

Run large language models locally on Intel Macs with AMD GPUs.

Native macOS app · Metal acceleration · No cloud, no accounts, no per-token costs

⬇ Download the latest release (opens in a new tab) · [ Changelog](CHANGELOG.md)

[Versión en español más abajo](#toshllm-en-español)

What is ToshLLM?

ToshLLM lets you run modern open LLMs entirely on your own Mac — your chats never leave the machine, there are no accounts, and there's nothing to pay per token.

Most local-LLM tools on macOS only target Apple Silicon. Intel Macs with discrete AMD GPUs — including Hackintosh builds — get left behind: the stock engines produce corrupted output on AMD dGPUs and read model weights over PCIe at a fraction of the possible speed.

ToshLLM fixes that. It bundles llama.cpp built with AMD-specific patches and wraps it in a polished native SwiftUI app — so a card like the RX 6700 XT goes from unusable to genuinely fast:

It opens, detects your hardware, and recommends models that will actually run well — no guesswork.

Features

• Native chat — multiple persistent conversations, full Markdown with code-copy, regenerate, system prompt, live tokens/sec, file attachments, conversation forking and per-message metrics
• Agent tools and MCP — the model can read, edit and run files and commands with per-step permission, run JavaScript in a sandbox, and use tools from external Model Context Protocol servers you connect
• Symbolic math (SymPy) — optional exact math tools for the model: simplify, solve equations and systems, derivatives, integrals, limits, series, differential equations, matrices, Laplace and Fourier transforms, and checking that two expressions are equivalent or that a solution holds. The runtime ships inside the app, nothing else to install; off by default, in chat settings under Agents ([details](helpers/tosh-sympy/README.md))
• Scientific computing (NumPy and SciPy) — optional numerical tools for the model: definite integrals without a closed form, roots, optimization and curve fitting, differential equations over an interval, linear algebra on decimal matrices, FFT and filters, and statistics. The model sends an operation and its numbers, never code. Same bundled runtime, its own switch, off by default ([details](helpers/tosh-scientific/README.md))
• **Math answers you can…