Oeon

Oeon is a personal intelligence that knows your work and gets it done.

Tell it what you need in plain words. Oeon plans the work, does it across your apps and files, and checks the result before you see it. It asks you before anything that can't be undone.

It remembers what you told it and what you decided, and forgets anything for good when you ask. Every result is checked before it counts, and every step is signed, so you can always see what it did. It runs on your own computer with Tin, our own model, so your work stays yours.

Ask it to reply to a customer, reconcile last month's payments, prepare the board report, find what was agreed three months ago, or forget a conversation for good.

Oeon

How it works

Oeon is a resident service on your own computer, on Windows, macOS and Linux. It needs no Oeon account. You connect a tool only when you want Oeon to use it.

Where it lives

In the background. Oeon runs as a quiet service on your computer. There is no window to keep open.
In your taskbar. Its icon sits by the clock, or in the menu bar on a Mac. A dot appears when something needs you. Click it to see what needs your approval and what was done today.
In your terminal. Type oeon to talk to it in plain words.

After you download

Install. Your invitation has one line to paste. On Windows, in PowerShell: irm https://oeonai.com/install.ps1 | iex. On a Mac, in Terminal: curl -fsSL https://oeonai.com/install.sh | sh. It downloads Oeon and Tin, checks them, and installs them as a program: nothing to unzip, no folder to manage, no Oeon account.
Choose what Oeon may use. Your files, your mail, your calendar: you pick each one. This is Oeon's mandate, and you can change it at any time.
See what it learned. After the first hour, the panel shows what Oeon now knows about your work and where it learned it. You can forget any of it, with a signed receipt.
Let it work. Oeon works in the background from then on, and asks you only before anything that can't be undone.

The loop

An event arrives. A mail, a ticket, a voice request, a schedule, or any program on the machine posting to Oeon's local API.
Tin decides. Your standing rules first, then Tin: ignore it, tell you, propose an action, or act. Tin never turns an event into an action on your files by itself.
The plan is fixed from your words. It is written before any untrusted text is read, and every value carries where it came from. A recipient that came from a mail, not from you, cannot receive a payment or a reply without your card.
Workers propose. Tin, or any engine you connect. Up to 100 run at once where a job splits into independent parts, under one conductor. They never talk to each other, only to the job's ledger.
The gate decides. A result counts only when a check outside the model passes: the project's tests on a fresh copy, a payment provider's receipt, a rule about where your data may go. A failed part goes back with the reason. It is never settled by a vote.
Irreversible acts wait. Paying, sending and deleting are held until the whole job verifies, inside the budget you set. Beyond it, you get a card. An uncertain payment blocks a second one until it is reconciled.
Everything is sealed. Each decision and act is signed with this machine's Ed25519 key into one hash chain, so an altered record fails verification. Tin learns only from outcomes a check confirmed.

What stays yours

Memory holds what you said, what Oeon verified and what you approved. A model's suggestion waits for your approval, and an instruction hidden in a memory is refused.
History keeps every past conversation on your machine, with secrets removed before anything is stored.
Forgetting is real. The text, its index entries and its place in the database file are erased, recall is checked, and the receipt is sealed.

Research

Tin is our own model: a decision layer and a world model, loaded together from one checkpoint. It answers with a probability for every option and never generates text. Everything below was measured on one laptop, an Intel Core i7-10510U with four cores and no GPU. Intervals are 95% bootstrap intervals on held-out rows.

Tin alone, beside the decision models

Tin's released checkpoint, on held-out rows of ten tasks rebuilt from LangWatch's sources.

Task	Tin	Jev	Kev-9B	Kev-4B	GLiNER2.5	Laya
Prompt injection	97.4	94.6	87.6	91.6	63.2	0
Personal data	99.8	90.8	62.5	9	16.8	15.8
RAG faithfulness	91.1	80.3	73.1	72.5	57.8	49.8
Banking77	91.9	79.6	85	85	70.6	n/a
Tool routing (BFCL)	89.4	78.3	72.4	71.4	32.4	22.9
Complaint routing	82.1	78.7	75	76.2	52.6	44.2
Commit type	68.7	68.3	55.5	53.7	50.4	44.8
Search relevance	60.5	57.7	43.4	44.8	25	31.5
Off-topic	89.3	93.4	89.6	84.4	49.3	54.3
Moderation (AUROC)	85.5	90.3	88.5	82	71.5	69.8

The best in each row is in bold. Read this as a reference, not a ranking: Tin was trained on each task's training rows, while the other models' figures are zero-shot on LangWatch's own test sets. On Kev's transfer suite, where every model answers the same 764 questions, Tin alone scores 61.3, against Jev 85.7, Kev-27B 84.8 and Kev-9B 82.2.

Banking77 intents

91.9% accuracy (90.2 to 93.6)

Complaint routing

82.1% accuracy (79.7 to 84.4)

Commit type

68.7% accuracy (65.6 to 71.5)

60708090100

The axis starts at 60, not 0.

Median time per decision: 0.2 to 2.6 milliseconds across ten tasks.
Calibration error: 0.01 to 0.08 across ten tasks. Lower is better.
Sure and wrong: On Banking77, Tin was at least 90% sure on 754 answers and wrong on 9.

Tin with a 9B open-weight model

A separate configuration for comparisons with larger models. The 9B model is an open-weight model used as published: we never train it, and it is not part of Tin. Tin answers first, and the 9B joins where Tin is least sure, with weights fitted on validation questions.

Benchmark (questions)	Tin with the 9B	Others, as published
Kev's transfer suite (764, the same for all)	82.6	Jev 85.7, Kev-27B 84.8, Kev-9B 82.2
Fast Decisions, single-label heads (166)	62.5	GLiNER2.5-Decide 60.2, GLiNER2 XL 59.6, JevK5 57.6, Laya Router 46.6 (their test split; Tin's is the development split)
BFCL function calling (50)	96	Jev 99.2, CLM-8B 95.2; the 9B alone also 96 on the same 50
Perplexity's decision panel: WinoGrande (50, preview)	72.0	pplx-decider-v1-27b 83.3, Jev 90.7 (on the full set)
Perplexity's decision panel: Belebele (50, preview)	84.0	pplx-decider-v1-27b 94.0, Jev 95.0 (on the full set)
MMLU-Pro (100, the same questions)	69	Claude Opus 5.5 96
HumanEval (40)	95	no published figure on the same problems
LiveCodeBench (10, diagnostic)	60	Claude Opus 5 89; the same 9B without Tin 50

World model: Before Oeon runs an action on your files, Tin predicts whether it will succeed, what it will create, change or remove, and whether it can be undone. On held-out episodes its success prediction was right 95.6% of the time. It advises; Oeon's gate keeps the final word.
When you have finished speaking: Over 300 spoken requests with a thinking pause inside, Tin cut the speaker off 1.3% of the time, against 20.3% for a fixed 1.7 second wait, with the same median wait.

What Tin does not do yet

It can't write, chat or summarise. It only chooses among the options it is given.
A new kind of decision needs training examples.
It is behind Jev on off-topic detection, moderation and Kev's transfer suite.
Its own speech model cannot transcribe yet: intent from speech on MINDS-14 is 12.5, against 83.7 for Oruk's model.
When a hard question goes to the 9B with thinking, an answer on a four-core laptop takes 15 to 35 minutes.
Not yet measured: GPQA Diamond, Humanity's Last Exam, OSWorld, Terminal-Bench and ARC-AGI. These run when Tin trains on a GPU.

Every figure, with its run report and caveats, is on Tin's model card: huggingface.co/Aravindhan11/tin.