We make small models from your own examples. Each one does your work your way, on your own device.
Your requests and its answers stay on your device.
Train your own AI. Run it on any device.
We optimize small AI models to run on any device, from your smartphone to your laptop, at good speeds. Yours learns your work from your own examples.
Join the waitlistJoining is free.
- Speed
- Your examples
- Meeting notes
- Emails
- Reports
On your device
Your modelSpeed
Your modelPreview
Built for any device.4-bit weights
From your smartphone to your laptop, at good speeds.
- Smartphones
- Tablets
- Laptops
Message your model
Your requests and its answers stay on your device.
Stop renting your AI. Own it.
Today the best AI runs in the cloud. You rent it, you hit usage caps, and every chat goes to the company that runs it.
No usage caps.
Ask as much as you want. You need no provider account.
No fee per message.
A model on your device has no price for each call.
It stays yours.
No company can change it or take it away.
It runs on your device.
Your chats stay there.
Trained on your data, for your workflow.
Show it your work.
Give it examples of the work, done the way you do it.
Summarise these meeting notes.
Decision: Launch moves to Friday. Owner: Sam. Next: Send the plan.
Answer in my style: “Can we move the call to Friday?”
Friday works for me. Same time?
Then it trains.
One step of training uses a cloud service. Your examples go to it if you agree, and not otherwise.
16.3 to 151.1 min
Training time for our task models on one computer.
Train it on the work you repeat.
- Summarise long reports.
- Draft emails in your voice.
- Turn notes into a plan.
- Explain a hard idea in plain words.
- Work through a problem step by step.
- Turn your notes into flashcards.
Faster than you can read, on any device.
A token is a small piece of a word.
Small from the start.
We begin with small open models, not giant ones.
Compressed.
We cut each weight to 4-bit precision, so the model fits in your device's memory.
Tuned for your hardware.
It runs on the graphics chip of your device.
How we measured speed
* Up to, from our trained Qwen3.5-0.8B model file qwen3.5-0.8b-Q4_K_M.gguf (Q4_K_M) on one tested computer with 16 GB. Reads and writes come from llama-bench on the GPU, the mean of 3 runs, a short diagnostic, not our ten-minute test.
Speed comes from the llama.cpp server on one tested computer, with one request at a time, a fixed prompt, and a fixed answer length.
For teams with files they cannot paste into a chatbot.
Give us examples of your team's work. We make a small model that does that work your way. You own the model, and you need no code.
Law firms, clinics, accountants, and banks.
One step of training uses a cloud service, if you agree.
One test item, two answers.
Order placed by C285-9029-478 on 2023-03-15. Total amount: $150.00. Shipping to 537 James Junctions, N4R9X4, South Sarahville. Contact bethany95@example.org or 2046606177 for inquiries.
Our model
- Customer ID
- C285-9029-478
- Street address
- 537 James Junctions, N4R9X4, South Sarahville
- bethany95@example.org
- Phone number
- 2046606177
Claude Opus 4.5
- Customer ID
- C285-9029-478
- Date
- 2023-03-15
- Extra field
- Street address
- 537 James Junctions
- Split across fields
- Postcode
- N4R9X4
- Split across fields
- City
- South Sarahville
- Split across fields
- bethany95@example.org
- Phone number
- 2046706177
- Wrong digit
Our model gave the expected answer, and Claude Opus 4.5 did not.
PILLOW DURIA 36,364 PLASTIK 25 0 SUBTOTAL 36,364 PD1 10% 10.000Z AMOUNT 3,636 TOTAL 40,000 CASH 50,000 CHANGE 10,000
Our model
- Subtotal
- 36,364
- Tax
- 3,636
- Total
- 40,000
- Cash paid
- 50,000
- Change
- 10,000
Claude Opus 4.5
- Subtotal
- 36,364
- Discount
- 10.000
- Extra field
- Tax
- 3,636
- Total
- 40,000
- Cash paid
- 50,000
- Change
- 10,000
Our model gave the expected answer, and Claude Opus 4.5 did not.
I have a duplicate charge
Our model
- Charged twice
Claude Opus 4.5
- Card payment declined
- Wrong value
Our model gave the expected answer, and Claude Opus 4.5 did not.
This BANKING77 score may be higher because some test messages closely match the examples used to train the local model.
On three tasks, a small model trained on examples rivals Claude Opus 4.5 with none.
Near-frontier on work like this. Big cloud models lead on long tasks with many steps. On many-step tool use, in a calibrated adaptation of BFCL, our 4B model scores 15.0 and Claude Opus 4.5 scores 81.0.
We trained a separate small model for each task, from Qwen3-1.7B.
Masks personal data
88.1%Our model95% interval: 85.9 to 90.2
70.7%Claude Opus 4.5
Personal details found and labelled. Misses and false finds lower the score.
Added after the results.
Sample of 300 synthetic documents.
- Claude Haiku 5.568.6%
Sorts banking questions
81.5%Our model95% interval: 79.0 to 83.8
73.9%Claude Opus 4.5
Share of questions put in the right group.
This BANKING77 score may be higher because some test messages closely match the examples used to train the local model.
On a pilot set of 300 items, a lookup with no model scored 66.0%, and our 0.8B model scored 79.3%.
- Claude Haiku 5.572.6%
Reads receipts
94.3%Our model95% interval: 91.9 to 96.4
89.9%Claude Opus 4.5
Fields read from each receipt. Misses and wrong values lower the score.
Added after the results.
Development result, 96 receipts.
The score leaves out 4 test receipts, because their answer keys hold lists.
Gap +4.4 points, 95% interval −0.1 to +9.3.
- Claude Haiku 5.560.3%
Full results
Both models answered the same test items, and the same code scored both. Claude Opus 4.5 received the task instructions and the allowed labels, data types, or field names, with no examples.
Everyday tasks. Mask personal data
Personal details found and labelled. Misses and false finds lower the score.
The band is the 95% interval.
- Our model: 85.9 to 90.2
- Claude Opus 4.5: 67.8 to 73.8
- Claude Haiku 5.5: 65.6 to 71.6
Gap in points
- Claude Opus 4.5+17.3+14.3 to +20.5
- Claude Haiku 5.5+19.5+16.3 to +22.6
Added after the results.
Sample of 300 synthetic documents.
Everyday tasks. BANKING77
Share of questions put in the right group.
The band is the 95% interval.
- Our model: 79.0 to 83.8
- Claude Opus 4.5: 71.1 to 76.5
- Claude Haiku 5.5: 69.8 to 75.3
Gap in points
- Claude Opus 4.5+7.6+4.5 to +10.7
- Claude Haiku 5.5+8.9+5.9 to +11.9
This BANKING77 score may be higher because some test messages closely match the examples used to train the local model.
Everyday tasks. Read receipts
Fields read from each receipt. Misses and wrong values lower the score.
The band is the 95% interval.
- Our model: 91.9 to 96.4
- Claude Opus 4.5: 85.3 to 93.9
- Claude Haiku 5.5: 51.9 to 68.6
Gap in points
- Claude Opus 4.5+4.4−0.1 to +9.3
- Claude Haiku 5.5+34.0+25.0 to +43.1
Added after the results.
Development result, 96 receipts.
The score leaves out 4 test receipts, because their answer keys hold lists.
Tool calls, BFCL V4. No call, real users
The band is the 95% interval.
- Our 4B model: 87.0 to 91.1
- Claude Opus 4.5: 81.0 to 85.9
Tool calls, BFCL V4. Call when a tool fits
The band is the 95% interval.
- Our 4B model: 50.5 to 89.8
- Claude Opus 4.5: 38.6 to 81.5
Tool calls, BFCL V4. No call when none fits
The band is the 95% interval.
- Our 4B model: 87.5 to 94.5
- Claude Opus 4.5: 80.9 to 89.7
On many-turn tasks we score 15.0 and Claude Opus 4.5 scores 81.0.
| One-turn calls | Real-user calls | No call, real users | Many-turn tasks, base set | Call when a tool fits | No call when none fits | Calls | No-calls | |
|---|---|---|---|---|---|---|---|---|
| Our 4B model | 82.6 | 69.9 | 89.3 | 15.0 | 75.0 | 91.7 | 93.8 | 70.0 |
| Its base, our 4-bit run | 82.4 | 73.3 | 82.8 | 23.0 | 87.5 | 90.8 | 94.6 | 46.7 |
| Its base, our 8-bit calibration run | 87.9 | 76.5 | 81.0 | 25.5 | 87.5 | 88.8 | ||
| Its base, BFCL's run | 87.9 | 76.4 | 81.1 | 27.0 | 87.5 | 88.8 | ||
| Claude Haiku 5.5 | No public BFCL row | 97.5 | 96.7 | |||||
| Claude Opus 5.5 | No public BFCL row | thinking cannot be turned off; its worst case at the reasoning caps breaks the Anthropic cap | ||||||
| Claude Opus 5 | No public BFCL row | 95.8 | 93.3 | |||||
| Claude Opus 4.6 | No public BFCL row | 98.3 | 96.7 | |||||
| Claude Opus 4.5 | 88.6 | 79.8 | 83.6 | 81.0 | 62.5 | 85.8 | 98.3 | 96.7 |
| MiMo-V2.6-Pro | No public BFCL row | 98.8 | 95.0 | |||||
| Kimi K2.6 | No public BFCL row | reasoning or thinking on a development item (32 reasoning tokens) | ||||||
| Qwen3-235B-A22B-Instruct-2507 | BFCL's FC run is not comparable: for this model BFCL renamed each tool name with a dot (a.b to a_b) before the call but graded each call against the name with the dot | 3 final failures in a row, last HTTP 429 | ||||||
| Gemma 4 26B A4B | No public BFCL row | 98.8 | 25.0 | |||||
| Kimi K3 | No public BFCL row | it always thinks, and its cost is high | ||||||
| GLM-5.3 | No public BFCL row | no off mode on DeepInfra, and its licence was not read | ||||||
How we tested privacy
Our 4B model ran 100 of 100 tool requests on one tested computer with direct network access and direct reads of the home folder denied, and with no API key in its environment.
On a reviewed sample, an estimated 65 to 76 percent of the holdout messages have a close match among the training examples (two reviewers, a strict rule). No independent negative control exists.
Our 4B model starts from Qwen3-4B-Instruct-2507. Mask personal data: Qwen3-1.7B. BANKING77: Qwen3-1.7B. Read receipts: Qwen3-1.7B.
Questions
The basics
What is Third Weather?
We optimize small AI models to run on your own device, from your smartphone to your laptop. Each model learns your work from your examples, and it is yours to keep.
What is a model?
The software that reads your request and writes the answer. A chatbot runs a very large model in a data centre. Ours is small enough for your own device.
How is this different from a chatbot?
A chatbot knows a lot about the world and nothing about how you work, until you tell it. Your model learns your tasks, your format, and your rules from your examples.
What can I train it to do?
Everyday knowledge work. Examples are summarising reports, drafting emails, explaining hard ideas, and working through a problem step by step.
How is Third Weather different from PocketPal, LM Studio, Ollama, and Unsloth?
You get a model made for your own work. PocketPal, LM Studio, and Ollama run models that other people made, and they work today. Unsloth gives you more control over training.
Cost and timing
What does it cost?
Joining the waitlist is free, and we ask for no card. We will publish the price before you pay anything. A model on your device has no fee per message.
When can I use it?
Join the waitlist, and we will tell you when our models are ready.
Your data
Where do my examples go during training?
You choose the examples. One step of training uses a cloud service. Your examples go to it if you agree, and not otherwise. We will name the service before you train a model.
What stays on my device?
When you use your model, your requests and its answers stay on your device. You need no provider account.
What do I own?
You get a model file, made from your examples. No company can change it or take it away. The licence of its base model still applies, and Licences lists it.
Does this page track me?
This page sets no cookies and loads no trackers. When the page loads, it requests files from this site only. To check, open the network tab of your browser.
Using it
How many examples do I need?
It depends on the task. Our task models trained on 581 to 3,826 examples. You will add examples in our app, or give them to us, with no code.
Will it make mistakes?
Yes. Every AI model makes mistakes. Check its work when the result matters.
Which devices does it run on?
Our early models come to macOS computers with 16 GB of memory. Then they come to computers with Windows or Linux, smartphones with iOS or Android, and tablets with iPadOS or Android.
Where do big cloud models still lead?
In long tasks with many steps. Results shows the score, and Full results shows each loss.
How did you test against Claude Opus 4.5?
Both models answered the same test items, and the same code scored both. Full results holds every score, interval, and note.