Development build of main (fb34fe6). It is not the latest release: sushi update, the built-in update check and Homebrew stay on v1.6.0 and never offer it. Ad-hoc signed and not notarized, like every sushi release.
To try it:
shasum -a 256 -c sushi-bin-macos-arm64.tar.gz.sha256
tar -xzf sushi-bin-macos-arm64.tar.gz
xattr -dr com.apple.quarantine sushi-macos-arm64
./sushi-macos-arm64/sushi --versionSince v1.7.0-dev3:
- Qwen3.8 runs on Sushi Cooler, the engine of our own, by default: its STQ packs and its BF16 checkpoint. Pass
--qwen-mlx-engine, or clear the web UI'sqwen_coolercheckbox, to keep it on MLX; the load line names the side. - A Qwen3.8 load on Sushi Cooler runs its vision tower there too, and decodes concurrent requests as joined rows, as on MLX.
- A Sushi Cooler load compiles nothing inside a request: its warm-up makes what a request can pick, and what a request still had to compile is kept for the next load.
- Sushi Cooler holds less memory per prefill chunk, and a prompt of several chunks reserves its KV cache once.
- DeepSeek-V4.1 on Sushi Cooler at
--ssd-budget-gb 108warms within the memory it has and counts its parked decode pool once. - Quality on Sushi Cooler (KLD to the first EOS, release-gate fixtures): Qwen STQ 2.75bpw 0.061793 (v1.6.0 on MLX 0.061834); DeepSeek STQ 1.75bpw 0.210682 (v1.6.0 0.210305).
Known in this build:
- A 4096-token Qwen3.8 prefill chunk on Sushi Cooler can hold about 19% more memory than it was admitted under.
- Builds of a single architecture (
-Darchs) without DeepSeek-V4.1 or without Qwen3.8 fail some tests; the release binary carries every architecture and is not affected.