GitHub radar
Splash: Local 27B AI Inference on Apple Silicon
Inco AI built an inference engine that runs 27B-parameter models on Apple Silicon Macs with no cloud required — through an API compatible with both Claude and OpenAI.
Inco AI built a local inference engine for Apple Silicon that runs 27B-parameter models entirely on-device, with no cloud required. It serves Qwen3.8-27B (17.4 GB) and Qwen3.6-35B (20.9 GB) through OpenAI- and Anthropic-compatible APIs, so coding agents connect without any reconfiguration. On an M5 Pro with 48 GB memory, it decodes Qwen3.8-27B at twice the speed of the next-fastest engine and achieves 282 ms time-to-first-token with a 32K cached context. Streaming, tool calls, JSON Schema constraints, images, and inline PDFs are all supported. Switching Claude Code, Codex, or OpenCode to the local engine is a single command — `splash claude`, `splash codex`, or `splash opencode`. Installation is via Homebrew, and the model server starts with one command.
Why a vibe-coder should care
If you have a high-end M-chip Mac and run routine agent tasks — formatting files, writing tests, light code review — you can point those tasks at a local 27B model instead of spending API credits. Be honest about quality: it is noticeably below frontier models like Opus 5 or Fable 5.1, so use it for repetitive, simple tasks where you already know the bar is low.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install Splash (local AI inference engine) on my Mac from https://github.com/incoai/splash — follow the README, install via Homebrew (`brew install incoai/tap/splash`), run `splash serve --model incoai/Qwen3.8-27B-Splash` and then `splash claude`, ask me if you need any permissions
Requires an Apple M3 or newer Mac with at least 36 GB unified memory — that means a Pro or Max chip, not a standard MacBook Air.
Open on GitHub▌ More finds