Its own dialect
Tool calls are taught and read in the format each model was trained on — XML, JSON, Gemma, LFM, K2, GLM, MiniCPM — taken from its chat template, not assumed.
Free and open sourcemacOS · Windows · Linux
The models on your disk,
put to work.
Chaty turns a model file into a whole workstation — chat, a coding agent, image and music studios, answers from your own documents, and a voice to talk to. It all runs on your Mac or PC. No account, no cloud, no telemetry.
Other platforms and what they need
Reasoning you can fold away; code, tables and math rendered as they stream. The agent found the upstream fix, patched the parser, added a test and ran the suite. An image model loaded: the app becomes a studio. Drawn on an M4 Pro. A music model loaded: a YuE2 song, rearranged on its own melody in a new style.
01Chat
A chat that shows its work.
Reasoning streams into a panel you can fold away. On models with an effort ladder — Qwen3.8, Muse-Glimmer, K2 Horizon — you choose how hard it thinks, message by message.
Answers render the way they were meant to: highlighted code, tables, KaTeX, Mermaid diagrams, and HTML you can run right there.
- What's on, at a glance
- A quiet line above the message box names what the next message will use — thinking, web search, the knowledge base — each one click from off.
- Pictures in, too
- Drop an image on any vision model — a GGUF with its projector, or an MLX build.
- Canvas
- Ask for a page and watch it build in a live preview beside its source; later changes land as edits, not rewrites.
- Web search, without a key
- Answers can pull in fresh results and cite them, through a chain of free providers.
- Every conversation, searchable
- Full-text search across your history; export any conversation to Markdown or JSON.
- Your way
- Warm or cool dark, paper or cream light, four code themes, English, 简体中文 and Português.
02Code
A coding agent for the model you can actually run.
Switch to Code, open a folder and describe the task. The agent reads, searches, edits and runs commands in your project — each step shown as it happens, each edit as a real diff.
When the turn ends, a card lists every file it changed. Open one to see the whole diff; undo one, or all of them.
- You stay in charge
- A live task plan. Edits and commands wait for your approval — or run on their own, if you allow it.
- Commands that keep running
- Dev servers and watchers move to the background; a command that asks a question gets a terminal to answer in.
- It remembers
- Project memory in plain Markdown. It can search its past sessions, and you can @-mention one to bring it in.
- Tools you add
- MCP servers (with a curated, tested list), skills written as Markdown, and a browser it can drive.
03Image
Pictures, drawn on your own GPU.
Load a text-to-image model and Chaty becomes an image studio: sessions that read like conversations, a live preview while it draws, and each model's recommended settings filled in for you.
It runs on stable-diffusion.cpp — Metal on a Mac, Vulkan on Windows and Linux — and on a Mac, MLX image models run on the MLX engine, faster still.
“A neon sign that reads LATE NIGHT DINER on a rainy street corner, reflections on wet pavement, cinematic”
Z-Image Turbo · Q4_K_M · 1024² · 8 steps · 2 min 51 s
“雨夜街角的霓虹灯招牌，写着「深夜食堂」，湿漉漉的路面倒映着灯光，电影感”
Qwen-Image 2.1 · Q4_K_M · 1024² · 20 steps · 16 min 10 s
Straight out of Chaty, unedited, on an Apple M4 Pro with 48 GB.
- The families that matter
- Z-Image and Z-Image Turbo, Qwen-Image 2.1, FLUX.1 dev and schnell, Chroma, Stable Diffusion 1.x, 2.x, XL and 3.x.
- The other files, found
- A text-to-image GGUF is only the denoiser. Chaty finds its VAE and text encoder beside it — or downloads the missing ones in one click.
- Keep going from a picture
- Start the next round from any image; models that edit take it as a reference and change only what you ask.
- Saved with its recipe
- PNGs carry the prompt and settings inside, filed by date — or save smaller JPGs.
04Music
Songs, sung on your own machine.
Load a text-to-music model and Chaty becomes a music studio: your style and lyrics as a turn in the conversation, the piece as a waveform to play, seek and select, and each family's recommended settings filled in for you.
It runs on audio.cpp in a helper of its own — Metal on a Mac, Vulkan on Windows and Linux. On an M4 Pro, YuE2 writes and sings a 30-second verse and chorus in 33 s; Stable Audio 3 Small makes 20 s in 1.2 s.
- Six families
- YuE2 (lyrics to a whole song, voice and band together), ACE-Step 1.5, MiniMax Music 3, HeartMuLa, Stable Audio 3 and MiDashengLM-Gen.
- Packages found whole
- A music model is often several files — a VAE, configs, a tokenizer, the parts of a larger model. Chaty finds them beside it, or downloads the missing ones in one click.
- Keep going from a piece
- Rearrange a song on its own melody or continue it from any second; repaint a stretch, cover it in a new style, or make a variation.
- Your lyrics, or none
- Write lyrics with section tags, leave them to the model, or ask for an instrumental; the style is a few words.
05Documents & research
Answers from your own documents.
Drop in PDFs, Word files, slides, spreadsheets, Markdown or code. Chaty indexes them on your machine, retrieves by meaning and by keyword, and cites the passage behind each claim — or says plainly when your documents don't cover the question.
When does the new pricing take effect, and what did finance approve?
The new pricing takes effect in Q3, once the security audit closes.1
Finance approved a budget of $1.2M for the rollout.2
Not covered by your documents: the hiring plan.
1roadmap-2026.pdf · p. 72finance-q2.md
- Scans, read page by page
- Scanned PDFs and photographed pages are transcribed by a vision model; charts inside documents can be described, so they're searchable too.
- Deep Research
- Several rounds of web search and reasoning, written up as a report that cites only what it used. Export to PDF or Markdown.
- Pick the sources
- Choose exactly which documents a question searches, and how many passages come back.
- A two-host podcast
- Turn a knowledge base into a spoken conversation between two voices, saved as a WAV.
06Voice
Or just talk to it.
Live mode is a hands-free conversation. Whisper listens, the model answers, and a local voice reads the reply back sentence by sentence — in English or Chinese. Cut a reply short and it is your turn at once.
Speech runs on the CPU, so it never competes with the model for GPU memory.
- Silence sends
- Stop talking, and your turn is sent — no button to hold.
- Read aloud, dictate
- Have any answer read to you, or speak a message instead of typing it.
- Interrupt
- Stop a reply mid-sentence and speak; the model remembers what it said up to there, not the rest it had written.
—Under the hood
Built around what small models get wrong.
A frontier model can hold a workflow together on its own. A model that fits on your laptop often can't: it repeats the call it just made, sends an empty argument, retypes the line it means to change slightly wrong, and calls the job done without running it. Chaty is engineered for that model.
Edits that land
An edit is checked against the file while it's still being written. A retyped line with the wrong spaces, escapes or line numbers still lands when exactly one place fits.
Slips caught at the step
Repeated calls, empty arguments and plans left unexecuted are noticed where they happen, with a correction that shows the model what it wrote.
Nothing done on red
A turn ends at a gate: whatever changed since the last passing run is verified first, and a red build is never reported as finished.
A cache that carries over
Each turn's prompt is built as an append to the last, so the model's cache is reused instead of read again from the top.
One app
No server to start, no port, no API key, no config file. llama.cpp on Metal or Vulkan; MLX natively on Apple Silicon.
—Models
The open models people actually use.
Any GGUF runs on llama.cpp; on Apple Silicon, MLX folders run natively too. These families get first-class handling — their templates, reasoning controls and tool-call formats wired in. Search and download from Hugging Face without leaving the app.
Language models
| Family | What's wired in |
|---|---|
| Qwen 3 · 3.5 · 3.6 · 3.8 | Thinking on or off; Qwen3.8's low · medium · xhigh effort ladder; vision where the model has it |
| Gemma 3 · 4 | Vision; Gemma 4's own tool-call format and reasoning |
| K2 Horizon · MoVA | Its effort ladder and arg_key tool calls, on both engines |
| GLM-4.5 · 4.6 · 4.7 | Its arg_key tool-call form, read and written |
| Llama 3 · Muse-Glimmer | Vision, and Muse-Glimmer's four-rung effort ladder |
| MiniCPM5 · LFM 2.5 | Each family's own tool-call format |
| Chaty · Qwen3.5-4B | Our own fine-tune for single-file web pages — one click at first launch |
Text to image
| Family |
|---|
| Z-Image · Turbo |
| Qwen-Image · 2.1 |
| FLUX.1 dev · schnell |
| Chroma |
| Stable Diffusion 1.x – 3.x |
Text to music
| Family |
|---|
| YuE2 |
| ACE-Step 1.5 |
| MiniMax Music 3 |
| HeartMuLa |
| Stable Audio 3 |
| MiDashengLM-Gen |
Community fine-tunes still chat: when a model's own chat template differs from llama.cpp's built-in guess, Chaty runs the template itself, as transformers would.
—Privacy
Private, because there's nowhere else for it to go.
Models, conversations, documents and pictures live in a folder on your disk. There's no account and no server of ours to trust — delete the folder, and it's gone.
Stays on your machine
- Everything you sayConversations, code sessions, voice.
- Everything you give itDocuments, their index, the pictures you attach.
- Everything it makesAnswers, edits, reports, images.
Uses the network — only as described
- Web search and Deep ResearchOnly when you turn them on.
- Downloads you startModels, voices, embedding files.
- One update checkA request to GitHub for the latest release, a few seconds after launch.
—Download
Bring the model home.
Free, open source, and yours to keep. Each button fetches the latest release.
Windows
- Needs
- Windows 10 or 11, x64
- GPU
- Vulkan — NVIDIA, AMD, Intel
- Package
- Installer, no admin needed
The macOS build is signed but not notarized, so the first launch needs one step. Clear the download quarantine once, then open Chaty as usual:
xattr -dr com.apple.quarantine /Applications/Chaty.app—Questions
Good to know.
Is Chaty really free?
Yes — free and open source, with no subscription and no account. You bring the models: search Hugging Face and download them inside the app, or point Chaty at a folder you already have.
Does it work offline?
Once a model is on your disk, yes. The network is used only for web search and Deep Research when you turn them on, for downloads you start, and for one check for a newer release after launch.
What hardware do I need?
An Apple-Silicon Mac, or a 64-bit Windows 10/11 or Linux PC. Small quantized language models run in 8 GB of memory; text-to-image models want 16 GB or more. Chaty sizes the GPU offload to your memory and refuses a model that cannot fit, rather than freezing the machine.
Which image models can it run?
GGUF builds of Z-Image and Z-Image Turbo, Qwen-Image 2.1, FLUX.1, Chroma and Stable Diffusion 1.x–3.x. A GGUF holds only the denoiser; Chaty picks up the VAE and text encoder from the same folder, or offers to download what's missing.
Which music models can it run?
audio.cpp's GGUF packages of YuE2, ACE-Step 1.5, MiniMax Music 3, HeartMuLa, Stable Audio 3 and MiDashengLM-Gen — the store's text-to-music shelf lists them. Pick a quantization and Chaty brings the package's other files with it.
On macOS it says the app “can't be verified.”
Chaty is signed but not notarized — there's no paid Apple Developer account behind it. Clear the quarantine once in Terminal:
xattr -dr com.apple.quarantine /Applications/Chaty.appOr open it, dismiss the warning, and choose System Settings → Privacy & Security → Open Anyway.
How is Code mode different from a harness pointed at Ollama or LM Studio?
A harness assumes a frontier model that can hold a workflow together on its own. A model on your own hardware often can't, so Chaty's agent is built around its habits: the prompt teaches the workflow, each slip is caught at the step it happens, and nothing is reported done on a red build. And it's one app — no server, no port, no API key.
Where is my data kept?
In a local SQLite database in your app-data folder, with pictures filed by date beside it. Settings → Data shows where, and can clear any part of it. Models can live on another drive — Settings → Model → Models folder.