mlx-ts

mlx-ts

Local LLM inference on macOS using a Swift MLX host process + a TypeScript client / AI SDK provider.

This README is shown on the npm package page. The repo contains additional development notes.

Quickstart (end users)

Requirements

• macOS Apple Silicon (darwin/arm64)
• Node.js

Install

npm i mlx-ts

During install, mlx-ts downloads a prebuilt mlx-host (Swift) binary + mlx.metallib from GitHub Releases (no Xcode required).

Use with the AI SDK

import { createMlxProvider } from "mlx-ts";
import { generateText, streamText } from "ai";

const modelId = "mlx-community/Llama-3.2-1B-Instruct-4bit";

const mlx = createMlxProvider({
  model: modelId,
  // optional:
  // modelsDir: "/tmp/mlx-ts-models",
  // hostPath: process.env.MLX_HOST_BIN,
});

const model = mlx.languageModel(modelId);

// stream
const s = await streamText({
  model,
  maxTokens: 64,
  messages: [{ role: "user", content: "Say hello from a local MLX model." }],
});
for await (const chunk of s.textStream) process.stdout.write(chunk);
process.stdout.write("\n");

// one-shot
const g = await generateText({
  model,
  maxTokens: 64,
  messages: [{ role: "user", content: "Summarize MLX in one sentence." }],
});
console.log(g.text);

Runtime configuration

• Force CPU vs GPU: set MLX_HOST_DEVICE=cpu (default is gpu).
• Override host binary: set MLX_HOST_BIN=/path/to/mlx-host or pass { hostPath } to createMlxProvider.
• Override where models are cached: pass { modelsDir } to createMlxProvider.
• Override where mlx-ts downloads assets from: set MLX_TS_HOST_BASE_URL (base URL containing mlx-host and mlx.metallib).

OpenCode integration

OpenCode supports OpenAI-compatible providers and allows setting options.baseURL (OpenCode Providers) and selecting models via provider_id/model_id (OpenCode Models).

mlx-ts ships a small OpenAI-compatible local server:

# Start local server (choose any MLX model id)
npx mlx-ts-opencode --model mlx-community/Llama-3.2-1B-Instruct-4bit --port 3755

# Generate an opencode.json snippet
npx mlx-ts-opencode --print-config --model mlx-community/Llama-3.2-1B-Instruct-4bit --port 3755 > opencode.json