Development build of main (a954450). 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-dev1, for DeepSeek-V4.1-Flash on Sushi Cooler:
- Sampled text no longer gets a random token slipped in. With neither top-p nor top-k set, about one token in 130 was drawn from the whole vocabulary whatever its probability (1.6.1-dev2 and 1.7.0-dev1), which broke tool calls in agent sessions. A seed's text changes once more with the fix.
- An agent turn no longer prefills the previous reply again: the next request starts from the state behind the reply it just generated, tool call included, and prefills only the tool results and what else is new.
- Parallel tool results are read in the order the calls were made, whatever order the client returns them in, and a call with no arguments is written as DeepSeek's own encoder writes it.
- No GPU kernel is compiled inside a request: replies that grew the context past 768, 1408 and 2048 tokens each paused once to compile one.
For every model: parallel tool calls sent through /v1/responses with no reasoning item before them are one assistant turn, as they are through chat completions. The server log names a turn that stopped because its tool call could not be read ([tool-parse]) and states each DeepSeek request's GPU time ([engine] GPU busy). The dashboard has an Expert streaming card.