New in MetalMind 1.3
Make your own model, on your own Mac
Fine-tuning used to mean a Python environment, a rented GPU, and a dataset you did not have. MetalMind turns it into a wizard: show a small model a handful of examples, and it learns to answer the way you want — your tone, your format, the role you gave it. Everything runs on this Mac, and nothing you show it is uploaded.
macOS 26 or later · Apple Silicon Mac
What a trained model is for
Four kinds of job, and they are all the same job underneath: you keep explaining the same thing in every prompt, and you would rather the model simply knew it.
Your style
Replies that sound like you, or like your company. The greeting, the length, the register, the sign-off — a model copies all of it.
A fixed format
Reports, summaries or tables that come out the same shape every time. Meeting notes into Summary, Decisions and Next steps, without restating the structure in every prompt.
A role
A tutor, a proofreader, a support agent — consistently, rather than for as long as a system prompt holds.
Sorting text
Labelling messages by topic, priority or tone, in the vocabulary your team actually uses rather than a generic one.
And what it is not for
It does not learn facts
Training teaches a model how to answer, not what is true. For questions about your prices, contracts or manuals, attach the files in a chat instead — then the model reads the real text every time.
It does not make a model smarter
A small model stays a small model. It gets better at your task, not better in general.
Your models only answer
They do not open files, search folders, or add calendar events. The regular models do that; a trained one is for replying.
Nothing is uploaded, ever
Your examples, the generated practice material, the training run and the finished model all stay on this Mac. There is no account to sign in to and no server to send them to.
The wizard asks you outright whether the model needs to know specific facts, and says the same thing if you answer yes. It is better to hear it on this page than after an evening of work.
What you need
A few examples
Three at the very least, six to ten is better. MetalMind writes the rest in the same style.
10 to 30 minutes
Most of it is waiting. You can close the window and it carries on; chat waits while the model is busy.
Power
Training works the chip hard. Plug the Mac in.
Pro, for the training step
Everything up to training — planning, examples, preparing the data — is free to try, and your prepared data is still there afterwards.
The wizard, step by step
Seven steps, and you can stop at any of them
Each step says what it is for, what a good answer looks like, and where to find what it asks for. Closing the window does not cancel anything — a running job carries on and the top bar shows it.
- 1 Goal
Say what the model should do
Pick the closest kind, then describe it in a sentence or two. Concrete beats broad: "Turn meeting notes into a report with Summary, Decisions and Next steps" is a goal; "be better at writing" is not. The wizard also asks outright whether the model needs to know specific facts — and if you say yes, it tells you that training is the wrong tool for that and why, before you have spent any time on it.
- 2 Examples
Show it how, a few times
Each example is a request and the exact reply you would want to it. Keep the style consistent and the topics varied: ten variations of one question teach one thing. Write them, import a file, or pull them out of conversations you have already had in MetalMind.
- 3 Model
Choose the model to teach
Only models that can actually be trained on this Mac are offered, each marked with its licence, its download size, whether it is a tight fit in memory, and a time estimate. Bigger models understand more and write better; they also take longer and need more memory. If your examples are not in English, the models that handle many languages are listed first — a small model that makes mistakes outside English passes those mistakes on to yours.
- 4 Data
Let it write the practice material, then check it
Your examples set the pattern and go in as they are. From them the model invents new requests and answers them in your style, up to the total you choose — fifty is a good start. Passages from documents you imported get a request written for them. Then the whole set is reviewed, problems are listed, and you untick anything that is wrong, odd, or not how you would have put it.
- 5 Train
Train it, and watch the distance fall
The model reads the examples again and again, adjusting itself slightly each time. A sixth of them is held back so progress is measured on questions it never practised: the curve going down means it is learning, and when it levels off or rises, training stops and keeps the best point. That is what stops a model from parroting your examples back at you. Close the window if you like — the top bar carries the progress and clicking it brings you back.
- 6 Compare
See what actually changed
The held-back questions go to the original model and to yours, side by side. If it is not right, the wizard says what to try: untick the examples in the wrong style, add better ones for the questions it got wrong, raise the total, or move up a model size.
- 7 Save
Keep it, name it, use it
Your model appears in the picker under Your models and chats like any other. It is the base model plus your adapter — a few dozen megabytes — so the original is untouched and still there. From the model list you can rename it, edit the instructions it runs under, train it again with more examples, export it, or go back to an earlier version.
Examples
You almost certainly already have the examples
The hardest part of fine-tuning is meant to be the dataset. In practice the material is sitting in your sent folder, in a spreadsheet, or in the chats you have already had — MetalMind reads JSONL, JSON, CSV and TSV, and takes PDF, Word and plain text documents as source material too.
Emails you have sent
A customer question and the reply you actually sent. Your best replies make the best examples.
A spreadsheet
Two columns — request and reply — saved as CSV. In Excel that is File › Save As › CSV. TSV works too.
Help-desk exports
Zendesk, Intercom, Freshdesk and most others export conversations as CSV or JSON. Import the file as it is; MetalMind accepts JSONL, JSON, CSV and TSV and works out the shape.
Documents you have written
PDF, Word or plain text. MetalMind works out what question each passage answers, and the model learns to write like them.
Your chats in MetalMind
Pick the conversations where the reply came out exactly right, and take the pairs straight from them.
Nothing at all
Then write them. Ten minutes of typing is plenty — you only need a handful, and private material is fine, because none of it is uploaded.
What makes an example good: exactly the reply you want, down to the greeting and the formatting, because the model copies all of it — including the mistakes. Varied requests rather than ten versions of one. Full replies rather than one word. And consistency: if two examples disagree about style, the model picks one at random.
Data quality
Checked before it trains, not explained afterwards
A bad training set does not fail loudly. It produces a model that is subtly wrong in a way nobody can account for twenty minutes later. So the set is reviewed first, and you are shown what is wrong with it while fixing it still costs nothing.
Duplicates and empty rows
Removed outright before anything is counted, so an export with repeated rows does not quietly weight one answer.
Contradicting answers
The same question with two different replies. The model would learn to pick one at random, so both are flagged.
Replies that echo the question
An artefact of most exports, and it teaches a model to restate rather than answer.
Examples too long to fit
Anything over the training cut-off would be truncated mid-reply. Split long texts into smaller pieces instead.
One-word answers
Measured, not assumed: hand-written one-word replies taught a 3B model to guess, getting three of eight facts right against seven when the fuller replies were used. If most of your replies are a few words, MetalMind says so.
Too few examples
Under ten there is no pattern to learn and the step will not continue. Forty or more is where results stopped looking like luck in the measurements.
Your models
What you are left holding
A few dozen megabytes, not a second copy
What you train is an adapter layered over the base model, like a lens over a camera. Ten variations of an idea cost ten adapters, not ten copies of the weights, and the original model is left exactly as it was.
Versions you can go back to
Rename a model, edit the instructions it chats under, or train it again with more examples. Earlier versions are kept, so a retrain that turns out worse is not a one-way door.
It is a standard adapter, not a lock-in
Export writes adapter_config.json and adapters.safetensors — the layout mlx-lm reads — with the command to run it. Import works the same way round, whether the adapter came from MetalMind or from your own mlx-lm run.
On the local API as custom/<name>
A trained model is served over the OpenAI-compatible server like any other, so Cursor, Zed or a script reaches it by changing the model string. This, and chatting with a trained model, is what Pro pays for.
What this costs
Choosing the goal, writing or importing the examples, generating the practice material and reviewing it are free, and your prepared data is kept. Training the model, chatting with the result, and serving it over the local API are part of MetalMind Pro — $4.99 for three months, $9.99 a year, or $16.99 once. The point of the split is that you decide with real examples in front of you instead of a feature list.
FAQ
Training a model, specifically
How many examples do I need to fine-tune a model?
What can I import examples from?
How long does training take on a Mac?
Which models can MetalMind train?
Can I take a model I trained to another tool?
Can a model I trained edit files or use my calendar?
Do I need to be online to train a model?
Everything else about MetalMind — what stays on your Mac, the local API, measured throughput, and the rest of what is free — is on the main page.