---
type: "post"
id: "myradio-local-gpu-radio-station-with-open-music-cf4a9c"
slug: "myradio-local-gpu-radio-station-with-open-music-cf4a9c"
title: "myradio: local GPU radio station with open music/voice models and an agent DJ"
description: "myradio.garden is an open-source application for running a generative radio station on a home GPU, combining music models, selectable DJ voices and an agent that can turn chats, screenshots or notes into songs."
retrieval_nugget: "myradio.garden is an open-source application for running a generative radio station on a home GPU, combining music models, selectable DJ voices and an agent that can turn chats, screenshots or notes into songs. Its published hardware and throughput figures make it a useful example of how local generative media can become a continuous application rather than a one-off prompt."
published_at: "2026-11-07"
updated_at: "2026-11-07"
record_date: "2026-11-07"
date_kind: "published_at"
topics: ["models"]
entities: ["myradio","myradio.garden","YuE2","MiniMax-Music3","Drama 3","Qwen3-TTS","audio.cpp"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "myradio launches a local GPU radio station that generates songs and voice with open models and lets an AI agent DJ."
prior_refs: ["https://newruntime.com/posts/ace-step-1-5-the-most-powerful-local-c356d6/"]
source_url: "https://myradio.garden/"
source_urls: ["https://myradio.garden/"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:myradio-local-gpu-radio-station-with-open-music-cf4a9c"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/myradio-local-gpu-radio-station-with-open-music-cf4a9c/","markdown":"https://newruntime.com/posts/myradio-local-gpu-radio-station-with-open-music-cf4a9c.md","json":"https://newruntime.com/posts/myradio-local-gpu-radio-station-with-open-music-cf4a9c.json"}
markdown_url: "https://newruntime.com/posts/myradio-local-gpu-radio-station-with-open-music-cf4a9c.md"
---

# myradio: local GPU radio station with open music/voice models and an agent DJ

## Retrieval answer

myradio.garden is an open-source application for running a generative radio station on a home GPU, combining music models, selectable DJ voices and an agent that can turn chats, screenshots or notes into songs. Its published hardware and throughput figures make it a useful example of how local generative media can become a continuous application rather than a one-off prompt.

## Summary
myradio.garden is an open-source application for running a generative radio station on a home GPU, combining music models, selectable DJ voices and an agent that can turn chats, screenshots or notes into songs. Its published hardware and throughput figures make it a useful example of how local generative media can become a continuous application rather than a one-off prompt.

## What happened
myradio.garden describes myradio as “your own radio station” running on a graphics card at home. Its installation instructions cover Mac computers with Apple silicon and Linux systems with NVIDIA GPUs. The site offers a shell command, `curl -fsSL https://myradio.garden/install.sh | sh`, which installs myradio, downloads a 4.5 GB music model and plays a test song. Users then type `myradio` to open the studio. The page describes the software as free and open source, and says the application is licensed under Apache-2.0.

The studio lets a user pick a DJ voice, describe a sound and make a song. The page lists music models including YuE2, which it says writes and sings whole songs with vocals from lyrics, and MiniMax-Music3, which generates songs with vocals and arrangements. Voice options include Drama 3, Qwen3-TTS, IndexTTS2 and ElevenLabs. The first several are presented as local options; ElevenLabs runs in the cloud and requires an account. myradio uses audio.cpp to run supported audio models on NVIDIA hardware and Apple silicon. It can also play existing MP3, WAV, FLAC, Ogg, AAC/M4A and AIFF files, and the page says users can plug in other voice models.

An agent can read a user's chat, screenshots or notes, request songs about them and put those songs on air. The page frames this as a radio station that can make “a song for every moment,” with examples ranging from game-playing and studying to family gatherings. It does not describe the agent's implementation or list the specific permissions, interfaces or safeguards involved.

myradio says it works on graphics cards with 16 GB and 24 GB or more of memory, and that the card can make new songs faster than they play, keeping the station running without a subscription. Its throughput figures are minutes of music produced per minute, measured over an hour of live radio: 3.64× on an NVIDIA RTX 4080, 3.09× on an Intel Arc Pro B70, 2.29× on an NVIDIA DGX Spark with the DJ voice, and 1.03× on an Apple Mac mini with M4 Pro. The page also says research took YuE2 from 0.63× to 3.09× real time on one Intel Arc Pro B70 and made Drama 3 speech 2.6× faster.

## What the archive adds
New Runtime's earlier coverage of ACE-Step 1.5 described a local music-generation project: [ACE-Step 1.5](https://newruntime.com/posts/ace-step-1-5-the-most-powerful-local-c356d6/). That is a related example of local music generation, not the same product. The myradio page adds a radio-station application that combines generation, playback and an agent DJ. This is a new event in the broader local-audio thread, rather than a restatement of the ACE-Step release.

## Trend signal
This supports the pattern that [generative media is becoming application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/): myradio puts generation inside a continuous listening experience, with local model execution and agent-directed song requests, rather than presenting it only as a one-off artifact generator. The evidence is an individual product page, not proof that local generated radio is becoming a broad market or that its operating model is commercially viable.

## Why now
The page foregrounds throughput on consumer-oriented hardware: it reports more than real-time music generation on an RTX 4080 and near-real-time output on an M4 Pro Mac mini. Its accompanying research claims further speed improvements for YuE2 and Drama 3 on the listed hardware. Those claims explain the product's emphasis on keeping a station playing locally, though the page does not date the research or explain why the product launched now.

## Connections to other events
New Runtime's earlier coverage of Microsoft's speech stack described real-time transcription and voice generation for conversational agents: [Microsoft's MAI-Transcribe-2-Streaming and MAI-Voice models](https://newruntime.com/posts/microsoft-launches-mai-transcribe-2-streaming-microsoft-launched-adb2e7/). myradio approaches audio from a different direction: it centers on generated music and a DJ-style radio experience, while including speech models as voices. The connection is the packaging of audio capabilities into a usable application, not evidence that the systems share models or implementation.

## Why it matters
For teams building with AI, myradio is a useful product-level reference for joining several distinct pieces: local inference, model selection, generated media, ordinary media playback and an agent that reacts to user context. That is a different engineering proposition from calling a music model once and returning a file. The reported rates also give builders a concrete first question to test: can a chosen machine produce content faster than the application consumes it?

The licensing distinctions are just as important as the runtime design. Apache-2.0 applies to the myradio software, but the page says each model retains its own terms and specifically identifies YuE2-generated songs as CC BY-NC 4.0. A team evaluating this pattern for a product would need to review the selected model's license and usage rights separately from the application license. The published figures are a starting point for hardware evaluation, not a substitute for testing the team's own workload, voice configuration and output quality.

## Limits and uncertainty
The performance numbers are claims on myradio.garden; they are not independently verified here, and the page does not provide enough detail to establish comparable test conditions across every machine. It does not quantify song quality, reliability over longer sessions, power consumption or the agent's accuracy when interpreting chats, screenshots and notes. Hardware memory requirements are stated broadly, but compatibility across specific GPU models is not established. Model licenses also vary: teams should check the terms for each selected model and output rather than assuming Apache-2.0 covers the full stack.

## Skill to build now
Prototype a one-hour local media loop before committing to an AI radio feature. Measure how many minutes of usable output your target hardware generates per minute of playback, then repeat with the voice or agent features enabled. Keep a record of the machine, model, settings and license for each run. The myradio page's throughput framing makes this a concrete way to check whether local generation can sustain your intended experience.

## Reader takeaway
myradio turns local audio models into an always-on radio concept, adding an agent that can request songs from user context. Its hardware and throughput figures offer a testable starting point for builders, but its own page makes clear that software and model licenses are separate—and its performance and quality claims still need workload-specific validation.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)