Get a copy
For creators and builders · runs on your hardware

Create with open AI.
Make it your own.

Make images, music, voiceovers, sound effects, voice swaps and short videos just by describing what you want, or build an AI model tailored to your needs. Your own copy, on your own machine, without a hosted provider's request quotas.

Get a copy See how easy it is

No coding needed to create: everything happens in your browser.

AI Studio — Audio Generation
The AI Studio Audio Generation page: Voice, Music and Sound effects tabs, a music style prompt, and a library with a voice swap, a sound effect, a voiceover and a 30-second song, each with Play and MP3 download buttons.

The real Audio Generation page in AI Studio, showing tracks it generated.

Easy to use

Describe it. Generate it. Use it.

Creating in AI Studio works like filling in a short form: no code, no command line, everything in your browser.

  1. Step 1: Pick what to make

    An image, a voiceover, a song, a sound effect, a voice swap or a short video. Each has its own page with the options laid out in front of you.

  2. Step 2: Describe it in plain words

    Type what you want, paste lyrics or a script, or upload a recording. For voices, you can record yourself right in the page.

  3. Step 3: Play, download, use

    Watch the progress while it works. Your result lands in your library with a built-in player, ready to download as an MP3, image or MP4.

  • No coding to create
  • Works in your browser
  • Record your voice in the page
  • Everything saved in your library

Installing AI Studio is a one-time setup, with step-by-step guides. After that, creating is point and click.

Why your own copy

Open models, your models, your rules

Hosted AI is quick to start with, but the model, the limits and the data handling belong to someone else. Running your own copy puts those decisions back with you.

Start from open models

Pull open text models through Ollama and try them in a chat playground. Images, speech, music, sound effects and video run on open models too.

  • Add, remove and switch models without changing providers
  • Ground answers in your own documents, with cited sources
  • Compare models with automated evaluation

Create a model that's yours

Turn your documents, a written goal or a codebase into a reviewed training dataset, then fine-tune an open model on it.

  • Approve, edit or reject every example before training
  • LoRA fine-tuning, with the result added to your models automatically
  • Or bake a system prompt into a base model when that's enough

Control over usage and deployment

The platform doesn't meter or cap your requests. How much you generate depends on your hardware, and you decide where it runs and who can use it.

  • Runs on your own machine or server
  • Your documents, datasets and outputs stay on it
  • Roles, revocable API keys and an audit log for teams
How it works: your own model

From an open model to your own, in three steps

Want an AI that speaks your language? Each step is a page in the app, with a review step wherever your data is involved.

  1. Step 1: Choose a model

    • Pull and manage open models from the Model Registry
    • Try them in the conversation playground, optionally grounded in your documents
    • Score them against your examples with automated evaluation

  2. Step 2: Make it yours

    • Generate training examples from documents, a goal, or a codebase
    • Review every example before it's used
    • Fine-tune with LoRA: MLX on Apple Silicon, Unsloth on NVIDIA GPUs

  3. Step 3: Generate and integrate

    • Chat with your model, with retrieval over your knowledge base
    • Call it from your apps through OpenAI-compatible endpoints and API keys
    • Send good conversations back into the dataset to keep improving it
For developers

Call your models from your own apps

Everything you build is available through OpenAI-compatible endpoints for chat, model listing and image generation, secured with API keys you issue and revoke yourself. Existing SDKs and tools can point at your own server.

your-server — terminal
Fine-tune complete · added to your models as support-assistant:v1
$ curl http://localhost:8000/v1/chat/completions \
    -H "Authorization: Bearer jz_live_…" \
    -d '{"model": "support-assistant:v1",
           "messages": [{"role": "user",
                         "content": "How do I reset my router?"}]}'
Open base model Your dataset LoRA fine-tune Your API

Illustrative example: calling a model you fine-tuned through the OpenAI-compatible API on your own server.

What you can make

Images, audio and video for everyone, and much more

Whether you're making a song, a voiceover, an image or a short video, swapping the voice on a track, or building an assistant for your business, it all happens in one studio, on open models running on your own hardware.

  • Images, in your own style

    Generate images from text with Stable Diffusion 1.5 or SDXL. Upload your own pictures, which are captioned automatically for review, and train a LoRA that learns a subject or style.

    Try“A cosy reading nook with warm afternoon light, watercolour style”

    Runs on: Generate on Mac or NVIDIATrain on NVIDIA

  • Voiceovers and narration

    Turn scripts of any length into natural speech. Build a voice from recordings of a speaker who has given consent, uploaded or recorded right in the browser. Every clip carries an inaudible watermark marking it as AI-generated.

    For exampleNarration for a training video, read in your presenter's voice.

    Runs on: Mac or NVIDIA GPU

  • Music and full songs

    Describe a style, add lyrics or keep it instrumental, and get a finished track as an MP3, from short loops to songs several minutes long. Train a style on music you own to make the results your own.

    Try“Warm lo-fi hip hop, mellow piano, vinyl crackle, 85 bpm”

    Runs on: Generate on Mac or NVIDIAStyle training on NVIDIA

  • Sound effects

    Create sound effects up to 30 seconds long from a plain description, for games, videos and apps. Fine-tune on your own sound library to match its character.

    Try“Thunder rolling in the distance with steady rain on a tin roof”

    Runs on: Generate on Mac or NVIDIAFine-tune on NVIDIA

  • Voice swap for songs

    Replace the singer in a song with a voice from a short recording. The melody, timing and words stay the same, and the new vocal is mixed back over the original music. Requires the voice owner's consent and rights to the song.

    For exampleHear your own demo song sung in your voice instead of a guide vocal.

    Runs on: Mac or NVIDIA GPU

  • Short video clips

    Generate clips of up to five seconds from a text prompt with Wan 2.2, at 480p or 720p, in landscape or portrait, saved as MP4.

    Try“A paper boat drifting down a rain-soaked street, cinematic close-up”

    Runs on: NVIDIA GPU

  • Assistants that know your documents

    Upload PDFs, Word files, web pages or spreadsheets, and chat with a model that searches them first and shows the sources behind each answer.

    For exampleYour product manuals, answering customer questions with the page each answer came from.

    Runs on: Mac, Linux or Windows

  • Specialist models, fine-tuned on your data

    Generate training examples from your documents or from a goal you describe, expand them into harder and edge-case variants, review them, and fine-tune an open model with LoRA.

    For exampleA model that answers in your support team's tone, or writes short bedtime stories for 5 to 8 year olds.

    Runs on: Apple Silicon or NVIDIA GPU

  • A model that understands your codebase

    Import one or more source projects as a zip. Training examples are built from each file together with the files it imports, so the model learns how your modules fit together.

    For exampleAn assistant that explains what breaks if an interface in your service changes.

    Runs on: Apple Silicon or NVIDIA GPU

And keep improving what you build

Score models automatically for accuracy, groundedness and hallucinations, keep versioned system prompts in a shared library, and send good conversations back into your training data for the next round.

Hardware notes show what each feature is built for; "Mac" means a Mac with Apple Silicon. The default speech, music, sound-effect and video models were chosen for licences that allow commercial use. Text and image base models keep their own licences, so check each one before commercial use.

Who it's for

For everyone who wants to create with AI

Creators and makers

Musicians, podcasters, video makers, designers, marketers and hobbyists: describe what you want and get images, music, voiceovers, sound effects, voice swaps and short videos, with no coding.

Developers

An OpenAI-compatible API in front of models you control, so existing SDKs and tools can point at your own server instead of a hosted one.

Startups and teams

Shared projects with owner, admin, editor and viewer roles, revocable API keys, and an audit log of every change.

Researchers and learners

Work through every stage of making a model, from dataset to fine-tune to evaluation, on hardware you can inspect and experiment with.

Many users, one studio

Bring your whole team, or your whole community

One installation of AI Studio can serve many people. Give each member their own sign-in and role, or open your studio to outside users who pay with credits.

Members and roles

Run one studio for your company, school, club or creative group.

  • Separate spaces. Each organization gets its own space, and members only see their own organization's projects and creations.
  • A role for every member. Owners and admins add members and choose what each one can do: OwnerAdminEditorViewer
  • Personal API keys that can be revoked at any time, plus an audit log of every change.
  • Fair sharing of one machine. Heavy image, audio and video jobs queue up and run one at a time, so a group can share a single GPU computer.

Offer it as a paid service

Share your studio with users outside your organization over the Nodofi network, and let them pay as they go.

  • Credit-based access. Users buy credits once, and each request uses a set number of credits by type: chat, image, voice, music, sound effect or video.
  • You set the prices, per credit and per type of request.
  • Automatic refunds. If a job fails, its credits are returned in full.
  • Get discovered. List your studio on the Nodofi Marketplace, with its models and prices shown.

Paid access runs through the optional Nodofi sharing service, with payments handled by MooChedda. Members and roles work on every installation.

FAQ

Practical questions

Do I need to know how to code?
Not to create. Making images, music, voiceovers, sound effects, voice swaps and videos, chatting with models, and building training data all happen in your browser by filling in simple forms. Installing AI Studio is a one-time technical setup with step-by-step guides, so you may want someone comfortable with that to set it up the first time.
Can many people use the same installation?
Yes. Each organization has its own members, roles and projects, and admins add members and set what each can do. Heavy generation jobs queue so a group can share one GPU machine. You can also offer the studio to outside users over the Nodofi network, with credit-based billing and automatic refunds for failed jobs.
What do I need to run it?
A Mac with Apple Silicon can run it natively, including LoRA fine-tuning of text models. Docker setups are included for Mac, Linux and Windows (through WSL2), plus a GPU setup for machines with an NVIDIA card. Image, music and sound-effect training, and practical video generation, need an NVIDIA GPU. Models download on first use, so plan for plenty of disk space: a few gigabytes per model, and tens of gigabytes for the largest.
How hard is the setup?
It's a real self-hosted application, not a one-click app. You run it with Docker Compose or install the parts natively, following the step-by-step installation guides that come with it. Once it's running, everything happens in your browser.
Which models can I use?
For text, any model you can run with Ollama. For images, Stable Diffusion 1.5 and SDXL models. Speech, music, sound effects and video use specific open models chosen for their licences: Chatterbox, ACE-Step 1.5, MOSS-SoundEffect, SoulX-Singer and Wan 2.2.
Do I own the models I create?
Your datasets, fine-tuned adapters and generated files are stored on your machine, under your control. Every base model keeps its own licence, and a fine-tune inherits the terms of the model it was built on, so check each base model's licence before using it commercially.
Are there usage limits?
The platform doesn't meter or cap your requests. The practical limits are your hardware's speed, memory and disk. GPU-heavy jobs such as image, audio and video generation run one at a time, so they don't run out of memory.
Does my data leave my machine?
Models run locally, and your documents, datasets and outputs are stored on the server you run it on. It needs an internet connection to download model weights.
Can I connect my own apps?
Yes. It has OpenAI-compatible endpoints for chat completions, model listing and image generation, secured with API keys that you issue and revoke yourself.

Create with AI, on your own terms

Make images, audio and video, swap voices on songs, or build a model of your own, all on hardware you choose, without a hosted provider's quotas in the way.

Get a copy Read the FAQ

Want to use AI Studio? Email aistudio@jozieai.com