Splash 1.0.1

Splash 1.0.1 improves memory handling, concurrent requests, and multimodal input support.

• More reliable startup and cache recovery under memory pressure.
• Shared-prefix reuse across concurrent requests and fairer prefill scheduling.
• Separate server ports with --port / SPLASH_PORT.
• Larger image uploads: 128 MiB by default, configurable with --max-request-size, with shared input budgeting and bounded upload time.
• PDF inputs up to 64 pages, subject to document, image, and context limits.
• Lower retained input/history overhead and a byte-bounded grammar source cache.
• Source-build compatibility for the placement-sparse capability query with older SDKs.

Upgrade

Stop running Splash servers, then:

brew update
brew upgrade incoai/tap/splash
splash --version

Restart with your usual splash serve --model owner/repo command. Existing model downloads and agent sessions are preserved. The prebuilt packages require no Xcode, Command Line Tools, or system Python.

Requires Apple GPU family 9 or newer and macOS 26.4+. Long cold prompts can still exceed a client's timeout; this update does not extend client deadlines.

Splash 1.0

Splash 1.0

Local inference on Apple silicon, accelerated by DFlash 2.

• Qwen3.8-27B and Qwen3.6-35B-A3B model packages.
• OpenAI Chat Completions and Responses, and Anthropic Messages APIs.
• Tool calling, structured output, images and inline PDFs.
• Launchers for Claude Code, Codex, OpenCode and Hermes.
• Precompiled Homebrew installation with a bundled Python runtime.

Requires Apple M3 or newer, macOS 26.4+, and at least 36 GB unified memory; 48 GB+ recommended. Available capacity depends on the model and workload.

brew install incoai/tap/splash
splash serve --model incoai/Qwen3.8-27B-Splash

Install your coding agent first, then connect from another terminal:

splash claude

Public model downloads require no Hugging Face token. Private or gated models require your own authorization.

Documentation · Launch post