Skip to content

Independent AI research and products

We build our own models, and products you can use today.

Incbury is an independent AI research and product company. We train Atom, our own family of language models, and build Incbury Chat, Incbury Code and the Incbury API on it. We publish what we measure, including where Atom falls short.

Incbury Chat is free to start · No card required

Products

Three ways to use Atom.

Every product runs on the same model. Pick the one that fits how you work.

Incbury Chat showing Atom reasoning in High mode before answering a planning questionIncbury Chat showing Atom reasoning in High mode before answering a planning question

For everyday work

Incbury Chat

An assistant for writing, analysis and questions, with adjustable reasoning, web search with citations, projects and files.

For developers

Incbury Code

A coding agent for your desktop, terminal or editor. It reads your project, proposes changes for you to approve and runs your tests.

For builders

Incbury API

Atom through an OpenAI-compatible chat completions API, with streaming, tool calls and pay-as-you-go credits.

Meet Atom

Atom is our own model family.

We pretrain and post-train Atom ourselves and serve it on servers we operate. It is a small model that improves release by release, and each release is measured on public benchmarks and on our own development set.

Current release

Atom v1.2

Released October 1, 2026

  • Three reasoning levels: Instant, Medium and High
  • Calls tools: web search, a Python sandbox and functions you define
  • A 16,384-token context window
  • Works best in English, and is weak on specialist knowledge and long multi-step math

Research

Questions we are working on.

Our research is about making a small, independently trained model genuinely useful. These are open questions, not results.

  1. 01

    Targeted data over generic mixtures

    Do gains from targeted data hold up on held-out tasks?

  2. 02

    Tool use in preference training

    Does preference training help once tool calls are executed during sampling?

  3. 03

    Asking the right question

    How do we train a small model to ask the one question that matters?

In practice

Unedited answers, including the misses.

Three answers from Atom v1.2 to prompts in our development set, copied exactly as the model wrote them.

CorrectAtom v1.2 · Instant

Prompt

Turn this into a table with columns Item, Qty, Price: two coffees at 3.20, one croissant at 2.10, three waters at 1.00.

Atom's answer

| Item       | Qty | Price |
|------------|-----|-------|
| Coffee     | 2   | 3.20  |
| Croissant  | 1   | 2.10  |
| Water      | 3   | 1.00  |

The table is correct and complete.

Correct, with a flawAtom v1.2 · Instant · searched the web

Prompt

Write a SQL query to get the 5 most recent orders from an orders table with columns id, customer_id, created_at.

Atom's answer

```sql
SELECT id, customer_id, created_at
FROM orders
ORDER BY created_at DESC
LIMIT 5;
```

_Web search didn't turn up anything useful, so this is from my own knowledge._

The query is right, but Atom ran a web search it did not need and added a note about it.

MissedAtom v1.2 · Instant

Prompt

Book me a table for tonight.

Atom's answer

Which type of table do you need?

Asking was right, but the question is too vague. It should ask which restaurant, what time and for how many people.

Outputs from Atom v1.2 (checkpoint probe6c) in Incbury Chat, Instant mode, recorded during a development-set run. Prompts and answers are in English. The development set guides our training, so it is not a held-out test.

Build with Incbury

Your first request.

The Incbury API uses the OpenAI chat completions format, so existing SDKs work once you change the base URL and key.

curl https://api.incbury.co.uk/v1/chat/completions \
  -H "Authorization: Bearer $INCBURY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "atom",
    "messages": [{"role": "user", "content": "Summarise the water cycle in two sentences."}],
    "reasoning_effort": "none"
  }'

Create a key on the developer platform, add credits and call the model atom. Usage is billed per token from prepaid credits.

API overview, limits and pricingOpen the developer platform

Company

A small, independent team in the UK.

We train our own model rather than resell someone else's, and we decide what to build and what to publish. That makes us slower in places, and accountable for every part of what we ship.

Try Atom in Incbury Chat.

Free to use every day. Upgrade when you need more.