ppmlx is an open source MLX-native inference tool for Apple Silicon. It is built to run LLMs locally on a Mac, with the page describing it as faster than GGUF translation layers and emphasizing local use without API keys, cloud costs, or latency.
The CLI centers on five commands and a zero-config workflow. It includes a built-in model picker, supports one-command launch of a coding agent or plain chat, and can start in under 10 seconds from a cold start. It offers hot-swapping models without restarting, an LRU cache for frequently used models, auto-download of missing models on first use, and a streaming REPL with token statistics, timing, and slash commands. The REPL also supports switching models mid-session with /model.
Model handling is a large part of the tool. ppmlx provides 168+ models from a curated MLX registry, with human-friendly aliases and Apple Silicon optimized weights. It supports multi-select downloads, stores models locally in ~/.ppmlx/models, and accepts any HuggingFace model for conversion to MLX format and 4-bit quantization in one step. The page says this can produce models that are 69% smaller with minimal quality loss and that quantized models fit in unified memory.
It also exposes an OpenAI-compatible API that works as a drop-in replacement for OpenAI-compatible tools and SDKs by pointing to localhost:6767. The server supports Chat Completions, Responses API, and Anthropic Messages, along with streaming, tools, vision, and function calling in XML and JSON. The page lists compatibility with Claude Code, Codex, Open WebUI, LangChain, and LlamaIndex, and says it supports vision and embeddings through mlx-vlm. Installation is shown with uv tool install ppmlx. The project is MIT licensed and built by The Focus Company.
In the Inference & model serving space, ppmlx takes a focused approach. It focuses on running large language models efficiently and natively on Apple Silicon devices without cloud dependency. It is built as an open-source project for AI researchers and developers using Apple Silicon. ppmlx is open source under the MIT license. It ships for the web, the command line, API, and macOS, and it can be self-hosted.
ppmlx first shipped in 2026. The project is developed in the open on GitHub with 25 stars and 113 commits in the last 90 days. Among its 8 catalogued features are OpenAI-compatible API, model hot-swap, and Streaming REPL. It exposes integrations via a public API.
Summary written by a language model from the project’s public pages.
What PulseGate has recorded for this listing
Closest matches by what these projects do