Repository navigation
Splash 1.7.0 speeds up decode in coding agents, fits more conversations in cache, and lets long sessions keep decoding together under memory pressure.
Engine
- Faster decode. On a 64 GB M5 Pro, recorded pi and Codex sessions on Qwen3.8-27B ran at 1.46× median decode speed and 1.23× median end-to-end speed versus the 1.6.2 development build. Bonsai reached 1.33× decode speed with four concurrent requests. Splash now verifies draft trees and prompt-lookup candidates in one pass, tuned to your Mac.
- Smaller cached states and faster follow-ups. INT8 storage cuts a cached Qwen3.8-27B recurrent state from 152.6 to 46.9 MiB. Splash saves state through the last completed page of a reply so the next turn can skip prefill for that part of the conversation.
- Better concurrency under memory pressure. Splash can free idle cache space to start a waiting session while another keeps decoding. In a two-session Qwen3.8-27B test under memory pressure, the second session waited 40% less to start. Decode also keeps moving as KV spills to SSD. The SSD quota tracks the space cached pages use on APFS. Hourly and daily write totals are available in
/statusand/metrics. - Models with streamed experts load their weights once at startup. Neural Engine prefill calibration better accounts for the cost of sharing work with the GPU.
Server and Mac app
- Control tags quoted in messages and documents stay literal. A role marker in a tool result can no longer forge a conversation turn. With thinking off, constrained JSON starts with the value, including when the output budget is small.
- Serve over a UNIX socket with
--host unix:PATH, set reasoning effort across requests with--reasoning-effort, or supply a custom--chat-template-file. The Messages API honors the same sampling fields as Chat. - Each chat keeps its own draft and attachments. If a send fails, the message returns to that chat’s composer. A new chat appears in the list once the server accepts its first message.
Upgrade
Choose Splash → Check for Updates… or download the DMG. With Homebrew, run brew update, then brew upgrade splash for the CLI or brew upgrade --cask splash-app for the app.
Persistent caches from 1.6.x start cold once. The first model start also calibrates decode for your Mac. Use --cached-states exact to store recurrent states at their original precision and --verify-rows all for repeatable greedy or seeded output.
Thanks to @JayToltTech, @Ptico, @oniisanlabs, @zaiyang-zhang, @Flo5k5, @pcvelz, @linson007 and @dmkjr for their contributions, and to @avifenesh and @jetnet for their reports.
Detailed changes and benchmarks · Full changelog since 1.6.1