Repository navigation
Prebuilt Swiftlet for Apple Silicon Macs (M1 or later, macOS 14 or later): swiftlet, swiftlet-server and swiftlet-repack, built from main at 36085b4.
Install
In Terminal:
curl -L https://github.com/leonickson1/Swiftlet/releases/download/v0.1.0/swiftlet-v0.1.0-macos-arm64.tar.gz | tar xz
cd swiftlet-v0.1.0-macos-arm64
./swiftlet-repack --from-hf Leonickson/Qwen3.6-35B-A3B-qpack --output ~/models/qwen3.6-35b.qpack
./swiftlet chat ~/models/qwen3.6-35b.qpack "What is the capital of Spain?"
The 4-bit Qwen3.6-35B container is 18 GB, so you need about 20 GB free. The download resumes if it is interrupted. For the 80B, use Leonickson/Qwen3-Next-80B-A3B-qpack (42 GB).
To serve an OpenAI-compatible API on 127.0.0.1:
./swiftlet-server --model ~/models/qwen3.6-35b.qpack --port 8080
The binaries are not notarized. Downloading with curl as above avoids the macOS warning. If you downloaded the archive in a browser, run xattr -dr com.apple.quarantine . inside the folder first.
Changes
- Prefill is about 2.5 to 3 times faster: tiled batched GEMV kernels and 512-token chunks, still bitwise identical to token-by-token decoding (#63). On a 24 GB M5 MacBook Air, a 512-token prompt to the 35B takes 7.5 s instead of 19.7 s, and 11 s instead of 33 s on the 80B.
- An experimental router lookahead with background expert prefetch, off by default (#62).
SHA-256 of the archive: bf6d0eae15bc24f284e6d5248101619e575cc528cfbd441fdfe1e0254042b2ee