Download

Tempo9 runs on Apple Silicon Macs (M1 or later) with macOS 26 or later. GPU acceleration (Metal) currently needs an M5 or later; on earlier Apple Silicon the engine runs on the CPU instead.

Release status

Tempo9 1.3.1 is out (2026-11-04), a bug-fix release: prompts now render from a model's chat template exactly as transformers renders them (Unsloth's Qwen3.6 and Gemma 4 E2B/E4B prompts carried extra whitespace before). It builds on 1.3.0 (2026-11-03), in which a server started with a vision encoder also serves a page in the browser: ask about an image, or keep an eye on a camera, a screen or a video file with rules that a local vision model checks only when the picture changes. Qwen3-VL images now get their DeepStack features. It builds on 1.2.2 (2026-11-02), in which --tower auto fetches the vision encoder that matches the model and Qwen3.8-27B loads on a 24 GB Mac, and 1.2.0 (2026-10-27), which added Flint (tempo9 --hf thinkspread/Flint-Q6K-IQ3XXS-GGUF). A signed, notarized binary, a Homebrew tap and the Swift package. Release notes · Gemma 4 E2B and E4B: Tempo9 vs llama.cpp · Flint on Apple Silicon: rec and Tempo9 compared.

Install

brew install thinkspread/tap/tempo9

or, without Homebrew:

curl -fsSL https://raw.githubusercontent.com/thinkspread/tempo9/main/install.sh | bash

Run

tempo9 --hf unsloth/Qwen3.5-9B-GGUF:Q4_K_S

That downloads the model into the standard Hugging Face cache on first use (a copy another tool already pulled is reused) and starts the server. A GGUF you already have works too: tempo9 --gguf ~/models/your-model.gguf.

Then point an existing client at it:

# OpenAI-compatible      base_url = http://127.0.0.1:11435/v1
# Anthropic-compatible   base_url = http://127.0.0.1:11435

Tempo9 has been tested with Codex, Claude Code, OpenClaw, and other local agent platforms that let you set an OpenAI or Anthropic base URL.

Port 11435 sits next to Ollama's 11434, so both can run side by side.

What you actually install

One binary: the Metal kernels are compiled into it. Every dynamic library Tempo9 links is already on macOS — no Python, no runtime, no framework, and nothing from Homebrew.

Measured on the same machine with du and otool -L; Tempo9 1.3.1's binary downloads as a 4.1 MB archive. llama.cpp's llama-server is a 49 KB shim over eight sibling dylibs in its build directory, and it also links Homebrew's openssl@3 — so it needs Homebrew installed and breaks if that formula moves. Ollama 0.32.15 installs 80 MB, though it bundles model management that Tempo9 does not have. The point is not the megabytes: it is that the file is the program, so there is nothing else to keep in sync with it.

Models

Tempo9 loads GGUF files directly. On first use it builds and caches an execution graph under ~/.cache/tempo9/; later runs reuse it.

SDK

Building a Mac app? A Swift package wraps the engine as a binary dependency. Its source is open (Apache-2.0) on GitHub: github.com/thinkspread/tempo9. SwiftPM downloads the engine from the release; declare platforms: [.macOS("26.0")] in your package.

.package(url: "https://github.com/thinkspread/tempo9", from: "1.3.1")

import Tempo9

License

The engine is licensed under the Tempo9 Engine License, free for individuals and small businesses: free to use if you are an individual, or a company with fewer than 100 employees and under $1M in annual revenue. Above that, a commercial license is required.

The Swift SDK and the server CLI are open source under Apache-2.0: github.com/thinkspread/tempo9.

For a commercial licence, email me.