Import your work, and a near-frontier model learns how you do it. It runs on anything from your smartphone to your laptop.
Join the waitlistWhat it learned from
A new document
We are delighted to announce that our new pricing will take effect in March.
Your model
New pricing starts in March.
Writes up to
75tokens per second
Reads up to
1,263tokens per second
A stranger read my report and asks a hard question.
You took the call.
A stranger asks for a call.
You said no.
“We can offer a discount.”
You changed it to “We may offer a discount.”
Agent with connectors and memory
Your note says no calls with strangers. Should I decline?
Your model
Yes. They did the reading, so they get the call.
Agent with connectors and memory
Your manager's email is marked urgent, and the deadline is today. Start there.
Your model
The customer. “Urgent” can wait an hour. A worried customer cannot.
We wrote these answers. Your own model will answer you.
Too big.
The strongest models need a data centre, not a phone.
Too slow.
Put one on a laptop, and you wait for every word.
Not good enough.
Small models run fast, but they get hard things wrong.
We make near-frontier models smaller,
so they run fast on any device.
Before training
Dear Team Member, I trust you are doing well. We are writing to advise that the Leadership Foundations program, coordinated by Elena Fischer, will begin on 7 May. It is anticipated that participants will need to set aside roughly three hours per week. We would be grateful if you could review the attached syllabus by 1 May. Please let us know if any adjustments may be required.
After training
Leadership Foundations starts on 7 May. Elena Fischer coordinates the program. You need three hours per week. Review the syllabus by 1 May. Next: send Elena Fischer your adjustments.
Before training
Good morning, everyone. As we continue to prepare for our transition to the new facility, it would be greatly appreciated if you could join a briefing on the relocation plan. We are tentatively planning to meet on April 2 at 2:00 p.m., and Daniel Okafor will walk us through the timeline. Should this conflict with your schedule, please do let us know at your earliest convenience.
After training
We plan to meet on April 2 at 2:00 p.m. Daniel Okafor walks us through the relocation timeline. You should attend this briefing. Next: Confirm your attendance with Daniel Okafor.
Before training
We are pleased to share that, following careful review, the organization has approved a hiring plan for the coming period. As part of this effort, Priya Raman will lead recruitment for 12 new positions beginning March 3. We recognize this may prompt questions, and we would be glad to address them in the weeks ahead. Thank you for your patience and continued commitment as we move forward thoughtfully.
After training
We approved a hiring plan for the coming period. Priya Raman leads recruitment for 12 new positions starting March 3. We will address your questions in the weeks ahead. Next: Ask Priya Raman about the 12 positions.
628 example edits. 11.1 minutes of training. 40 new documents, cut from 67 words to 28 on average.
88.1%Our model
70.7%Claude Opus 4.5, no examples
81.5%Our model
73.9%Claude Opus 4.5, no examples
94.3%Our model
89.9%Claude Opus 4.5, no examples
Near-frontier on tasks like these. Big cloud models lead most on long tasks with many steps.
Benchmarks are Gretel PII, BANKING77, and CORD.
Our model, Claude Opus 4.5, and Claude Haiku 5.5 answered the same test items, and the same code scored them. The two Claude models received the task instructions and the allowed labels, data types, or field names, with no examples.
We trained a separate small model for each task, from Qwen3-1.7B. Our 4B model starts from Qwen3-4B-Instruct-2507.
Everyday tasks. Data privacy (Gretel PII)
Personal details found and labelled. Misses and false finds lower the score.
95% interval
Gap in points
The band is the 95% interval.
Added after the results.
Sample of 300 synthetic documents.
Everyday tasks. Everyday tasks (BANKING77)
Share of questions put in the right group.
95% interval
Gap in points
The band is the 95% interval.
This BANKING77 score may be higher because some test messages closely match the examples used to train the local model.
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.
On a pilot set of 300 items, a lookup with no model scored 66.0%, and our 0.8B model scored 79.3%.
Everyday tasks. Reading documents (CORD)
Fields read from each receipt. Misses and wrong values lower the score.
95% interval
Gap in points
The band is the 95% interval.
Added after the results.
Development result, 96 receipts.
The score leaves out 4 test receipts, because their answer keys hold lists.
95% interval
95% interval
95% interval
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 | FT-300 calls | FT-300 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.
Import your work.
It learns how you work.
It runs on your device.
Third Weather gives you a private local AI. We optimize 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.
Yes. Import your emails, docs, and past chats, and your model picks up where you left off.
An agent looks things up. It finds what someone wrote down, and each request sends that text to a model in the cloud. Training changes the model itself, so it carries your judgment and your way of working, with nothing to search. For facts that change, such as today’s calendar, a lookup is still the right tool.
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.
When you use your model, your requests and its answers stay on your device. You need no provider account.
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.
We will publish the price before you pay anything. The waitlist asks for no card. A model on your device has no fee per message.
Alpha testing is running now, and we will tell you when our models are ready. If you would like to be part of it, let us know at [email protected] or join the waitlist.
Our early models come to macOS computers with 8 GB, 12 GB, and 16 GB of memory. Then they come to computers with Windows or Linux, smartphones with iOS or Android, and tablets with iPadOS or Android. Initial testing was done on a Samsung S23 Ultra with 12 GB of memory, a 13-inch MacBook Pro with M1 and 8 GB of memory, and a 13-inch iPad Pro with M5 and 12 GB of memory.
Most of all in long tasks with many steps. Full results shows each score and each loss.
This page sets no cookies and loads no trackers. When the page loads, it requests files from this site and from no other. To check, open the network tab of your browser.