san-digital/tokspur

Private chat and code on your Mac.

Website | Get started | Docs | Engine

Tokspur runs Qwen3.8 Flash-Next locally on a Mac. It uses a native Metal worker and serves an OpenAI-compatible chat API.

Tokspur needs an Apple M5 Max with a 40-core GPU, 128 GB of unified memory and macOS 27 or later. Run the installer in Terminal, then follow the setup guide to verify the installation and connect a client.

curl -fsSL https://tokspur.com/install.sh | sh
export PATH="$HOME/.local/bin:$PATH"

Add the model

Use the pinned target shards and shared Q8 sidecar from the model guide. Adoption reuses the original files. Check available space before building the local pack.

tokspur adopt --target /path/to/UD-Q4_K_XL --draft /path/to/snapshot
tokspur configure
tokspur pack plan --check
tokspur pack build
tokspur serve

Connect to http://127.0.0.1:18081/v1 with model tokspur and the generated API key.

Build

Install Xcode with the Metal toolchain, Git, Python 3.12 and Rust 1.99.1. Run these commands from the checkout root.

python3.12 scripts/bootstrap.py
.venv/bin/tokspur doctor

Licences

Apache 2.0. See NOTICE for third-party credits and model licences for the weights. Report vulnerabilities through the Security page.