Positioning Swama is a local AI runtime built for Apple Silicon Macs by Trans-N-ai. It is written in Swift, sits on top of Apple's MLX framework through mlx-swift, is a native implementation, and it installs through Homebrew.
What it does The runtime can run language, vision, embedding, speech-recognition and text-to-speech models. Beyond generation, it exposes an OpenAI-compatible API, a command-line interface and a menu bar app. Decision questions are served on two endpoints: POST /v1/decisions, which follows the OpenAI Decisions API style with predicate, choice and score questions, and POST /v1/systemone, a SystemOne OpenAPI 0.2.0 adaptation layer that reuses the same decision scorer as the Decisions endpoint. The /v1/systemone route accepts an explicit local model, a text or structured state, and named questions, and the TypeSafe JavaScript SDK can call it once it points at the local base URL. Image input is accepted as base64 data URLs carrying PNG, JPEG or WebP data.
Characteristics Swama is MIT-licensed, and its README ships in English, Chinese and Japanese. The README notes that the confidence a model reports is a "concentration" rather than a calibrated correctness rate. As of 2026-10-09 the GitHub repository had 593 stars and 31 forks. It was created on 2025-06-04, was last pushed on 2026-10-07, and has 17 releases in total, with the latest, v2.5.1, published on 2026-10-07.
When to use Swama suits running local decision models and multimodal models on a Mac behind an OpenAI-compatible interface. It is a macOS-only runtime: because it supports Apple Silicon only (macOS 15.0+ per the repository README; the v2.5.1 release assets list 15.6+), it does not run on Intel Macs or other hardware.