beamivalice/sushi v1.5.1

sushi v1.5.1

on GitHub

v1.5.1 — Thinking effort settings, fuller prompt-cache reuse, faster Qwen trellis packs

• Thinking effort is a setting: Settings › Inference › Thinking effort sets, for Qwen, DeepSeek and GLM, which
  effort words are offered, what each one asks the model for and how many thinking tokens it may spend, from the next
  request with no reload. DeepSeek-V4.1-Flash gains real medium and xhigh levels, and Qwen set to think off stays
  off. --reasoning-budget is removed: use the table, or a request's own reasoning_budget_tokens / budget_tokens;
  a numeric reasoning_effort is now refused instead of ignored.
• Follow-up turns reuse the whole conversation: Qwen3.8-Flash-Next restores the previous reply on the next turn
  instead of reading it again, which halved the prompt work across our eight-category test conversations, in the RAM
  and the SSD prompt cache. A full memory cache now trims an idle session back to a restore point instead of dropping
  it, and DeepSeek-V4.1 restores an image conversation after a restart again. Thanks @cyrilmhansen for the report.
• Faster Qwen trellis packs, and DeepSeek prefill: Qwen3.8-Flash-Next packs with a trellis trunk decode code about
  15% faster with prompt lookup and about 6% faster overall, prefill long prompts up to 10% faster, and answer the
  first prompt after a load about twice as fast. DeepSeek-V4.1 prefills 2-3% faster, and its first long prompt after
  an update about 9% faster. Output is unchanged.
• A clearer web page and menu bar: the Dashboard has fourteen charts you can pick and order, which move only when
  something changes; the web page and sushibar share plain labels (Generation speed, Prompt speed, Prompt cache), and
  every chart, setting and badge explains itself in a tooltip that also names the engineering term. A page opened at
  a LAN address of a --host 0.0.0.0 server can now load and unload models. A Cancel button stops a running request
  (thanks @chuaaron), and /metrics reports speculative decoding and prompt cache tiers.
  Settings gains "Skip the memory check before loading" for a Mac whose model fits but is refused (thanks @Cianidos).
• Download packs from the web page: Models › Available lists the published Sushi packs, recommends one for this
  Mac and downloads it with progress, Stop and Resume. A pack that is being updated, or whose update stopped half way,
  cannot load until it is complete, and Settings can turn every automatic update check off.
• sushibar does more: Log, Update and Quit buttons, a status card that says why a server stopped, servers bound to
  a Tailscale or LAN address, a Listen on LAN switch for the servers it starts (thanks @corysus), and a Port menu
  that keeps the ports you use. The default port is a setting.
• MiMo-V2.6-Flash leaves the supported list: it still loads and serves, but releases no longer test, benchmark or
  tune it; 1.5.0 was its last supported release.