Two free LLMs came with your Mac.
apple-llm calls both.

apple-llm speak "Hello there"                     # out loud, in your System voice
apple-llm speak "Hello" --voice Samantha --speed 1.2
apple-llm speak --file book.txt -o book.m4a       # save it, any length
apple-llm voices                                  # the voices you have
apple-llm transcribe meeting.mp4                  # audio or video to text
apple-llm transcribe meeting.mp4 -o meeting.srt   # subtitles
apple-llm listen                                  # the microphone, live

import { listen, speak, synthesize, transcribe } from 'apple-llm';

await speak('Hello there', { voice: 'Samantha', speed: 1.2 });

const { audio } = await synthesize('Hello there', { format: 'wav' });
await synthesize(longArticle, { voice: 'Samantha', out: 'article.m4a' });

const { text, segments } = await transcribe('meeting.mp4');

const said = await listen();   // returns when you pause

from apple_llm import listen_once, speak, synthesize, transcribe

speak("Hello there", voice="Samantha", speed=1.2)

audio = synthesize("Hello there", format="wav").audio
synthesize(long_article, voice="Samantha", out="article.m4a")

result = transcribe("meeting.mp4")   # .text, .segments

said = listen_once().text            # returns when you pause

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:11436/v1", api_key="unused")

speech = client.audio.speech.create(model="tts-1", voice="Samantha", input="Hello there")
speech.write_to_file("hello.aac")

client.audio.transcriptions.create(model="whisper-1", file=open("hello.aac", "rb")).text

On apple-llm serve, from the npm package. macOS can't encode MP3, so a request for it gets AAC, with a Content-Type that says so.

import { AppleLLM, tool } from 'apple-llm';
import { z } from 'zod';

const llm = new AppleLLM(); // 'auto': the device when it can, else the cloud

// Guaranteed shape on device: the decoder cannot produce anything else
const Contact = z.object({ name: z.string(), email: z.string().email() });
const contact = await llm.json('Reach Grace at grace@navy.mil', { schema: Contact });

// Your own functions, called mid-generation
const getWeather = tool({
  name: 'getWeather',
  description: 'Current weather for a city',
  parameters: z.object({ city: z.string() }),
  execute: async ({ city }) => ({ city, tempC: 21, sky: 'sunny' }),
});
await llm.text('Do I need an umbrella in Paris?', { tools: [getWeather] });

// Stream; breaking out of the loop stops the model
for await (const delta of llm.stream('Write a haiku about bridges')) {
  process.stdout.write(delta);
}

from apple_llm import AppleLLM
from pydantic import BaseModel   # optional: pip install 'apple-llm[pydantic]'

class Person(BaseModel):
    name: str
    born: int

with AppleLLM(tier="device") as llm:
    person = llm.json("Ada Lovelace, born 1815.", schema=Person)
    print(person.born)  # 1815, a validated Person rather than a dict

    llm.text("Read the total.", images=["receipt.png"], tools=["ocr"])
    llm.stream("Count to three.", on_delta=print)

with AppleLLM(tier="cloud") as cloud:  # after: apple-llm setup-cloud
    cloud.text("Solve this step by step: ...")

npx apple-llm serve   # OpenAI-compatible API at http://127.0.0.1:11436/v1
# options: --port 11436  --api-key <key>  --cors <origin>  --tier cloud

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:11436/v1", api_key="unused")
client.chat.completions.create(
    model="apple-on-device",  # or "apple-private-cloud"
    messages=[{"role": "user", "content": "Tag this: sourdough with a rye starter"}],
)

Works with the official SDKs, LangChain, LlamaIndex, Open WebUI and editor plugins. It supports streaming, tools, JSON Schema response_format and images, and binds to localhost only.

import { generateText } from 'ai';
import { apple } from 'apple-llm/ai-sdk';

const { text } = await generateText({
  model: apple(), // on device; apple('cloud') for Private Cloud Compute
  prompt: 'Write a one-line release note for a bug fix.',
});

AI SDK 6 and 7: generateText, streamText, generateObject, multi-step tools, images and abort signals.

apple-llm chat                                     # interactive, streamed
apple-llm run --schema person.json "Ada, 1815"     # guaranteed JSON
apple-llm run --tool ocr --image receipt.png "Read the total."
apple-llm run --session trip "My cat is Biscuit."
apple-llm rewrite "gonna grab a bite"              # also proofread, summarize, draft
apple-llm ask-screen "Summarize what is on this schedule."
apple-llm count --system "You are terse." -        # tokens before sending
apple-llm speak "Hello there"                      # out loud
apple-llm transcribe meeting.mp4                   # audio or video to text

The pip and npm packages install the same apple-llm command; chat and serve come with npm.