Models

Atom, our own model family.

Atom is a family of mixture-of-experts language models that we pretrain and post-train ourselves. Every Incbury product runs on it. This page covers how it works, which version is serving and what it cannot yet do well.

Talk to Atom Evaluations

16+1

Experts per layer: two routed per token, plus one shared

64k

Token vocabulary

16k

Token context window

Architecture

A sparse mixture of experts.

Each layer holds 16 expert networks and one shared expert. A router sends each token to the two experts that suit it best, so only part of the network runs for any one token. Most layers attend over a sliding local window; every fourth layer attends globally.

Inside one Atom layer

illustration · 2 of 16 experts per token

Mixture of experts

Only the experts a token needs are run, which keeps responses quick on modest hardware.

Three thinking levels

Instant, Medium and High trade speed for depth on each message.

Tool use

Atom is trained to call web search, a Python sandbox and developer-defined functions, and to cite what it finds.

Reasoning modes

Choose how much thinking each message deserves.

Every message in Incbury Chat can be sent at one of three thinking levels. Atom shows its reasoning as it works, then gives you the answer.

Fastest

Instant

Replies begin immediately. Ideal for drafting, rewriting, quick facts and back-and-forth conversation.

Balanced

Medium

Atom pauses to reason before answering. Better for analysis, planning and multi-step questions.

Most thorough

High

Extended thinking for hard problems: maths, code, research synthesis, with the reasoning shown as it happens.

More about reasoning modes

Versions

Release history.

Atom v1.2 serves every Incbury product today. Each release links to its report where one exists.

Atom releases
Version	Checkpoint	Released	Status
Atom v1.2	probe6c	1 October 2026	Serving	Report
Atom v1.1	if-sft	16 September 2026	Retired
Atom v1.0	Not recorded	10 September 2026	Retired

Public benchmarks for Atom v1.1

Atom v1.2 has not yet had a public benchmark run, so the most recent results are for Atom v1.1. Development-set results for Atom v1.2 are in its release report.

Benchmark results with the settings used
Release	Suite	Dataset	n	Mode	Decoding	Tools	Score
Atom v1.1if-sft	Instruction following	IFEval	541	Instant	greedy	No	41.8% prompt-level / 52.0% instruction-level (strict)
Atom v1.1if-sft	Maths reasoning	GSM8K test, 300-problem subset	300	Medium	greedy	No	27.0%
Atom v1.1if-sft	Code	HumanEval	164	Instant	greedy, 512-token limit	No	12.2% pass@1
Atom v1.1base	Knowledge	MMLU, 5-shot	14,042	Base model	log-likelihood	No	27.2%
Atom v1.1base	Commonsense and language	HellaSwag / ARC-Easy / ARC-Challenge / PIQA / Winogrande	16,695	Base model	log-likelihood	No	55.5 / 63.3 / 34.6 / 73.5 / 57.6

All evaluations Atom v1.2 release report

Limitations

What Atom cannot yet do well.

Atom is a small model. We would rather you knew its limits before relying on it.

Its general knowledge is limited. Atom v1.1 scored 27.2% on MMLU, close to chance on four-option questions. Turn on search for facts that matter.
Multi-step maths is unreliable without tools. Atom v1.1 solved 27.0% of a 300-problem GSM8K sample at Medium effort, without the Python sandbox.
Strict formatting instructions are often missed. Atom v1.1 followed 41.8% of IFEval prompts in full.
Its clarifying questions can be vague, and it sometimes searches the web when it does not need to.
It works best in English.
It can be confidently wrong. Check anything important, especially when no source is cited.

How we build Atom