Pienomial Launches AT0M for Business Decisions on Companies’ Own Hardware

Pienomial launched AT0M on September 30, 2026, offering businesses a model for routine decisions rather than open-ended writing. The company says customers can own it, train it and run it on existing hardware while keeping data inside their systems—a pitch aimed at organisations that want automation without sending those decisions to a hosted model.

A narrower job than a chatbot

The launch starts with a specific view of business work: many decisions are small, clearly defined and repeated. Pienomial is targeting those choices, not presenting AT0M as an assistant for unrestricted conversation. CEO Sanat Mohanty framed the product as an alternative to using a large hosted model for jobs that need a reliable choice within an established workflow.

Most business decisions are small, well-defined, and repeated thousands of times a day. They don't need a giant hosted model. They need something reliable that you own and control. That's what AT0M is,

Sanat Mohanty, CEO of Pienomial, in the company’s release

AT0M’s response to that problem is a constrained output: it selects from options developers define rather than generating free text. Pienomial calls this “System One” AI, meaning structured decisions for software. The practical distinction is the answer format. A developer supplies the possible responses, and the model chooses within that set instead of composing its own wording.

The decisions Pienomial is targeting

• Routing a request to the appropriate team.
• Determining whether an alert needs attention.
• Assessing how severe a reported problem is.

The company says routine cases can be automated while uncertain or high-stakes decisions go to a person for review. That leaves room for human judgment without requiring every request to receive manual attention. But a permitted answer is not necessarily a correct answer: restricting the menu controls the form of the response, not whether the model picked the right option.

Ownership reaches into training and deployment

Pienomial says organisations own AT0M outright, with no per-seat or per-token charges. The promise covers both training and operation within the customer’s environment, without requiring data to leave its systems. Those are distinct parts of the offering: ownership describes the commercial arrangement, while local training and execution describe where the organisation’s data is used.

The company describes AT0M as a single executable with no network dependencies. It lists support for Intel Xeon, Apple Metal and NVIDIA CUDA, spanning several computing platforms rather than one specified machine. Pienomial identifies banks, government agencies, healthcare organisations and life sciences companies as potential users where data must remain on premises. Those are intended customers, not announced deployments.

The first performance evidence comes from Pienomial

In company-run testing on the public typed-decisions benchmark, Pienomial says AT0M matched reference answers on 78.9% of 2,000 decisions. Its release cited a reported 74% score for TypeSafe’s Jev. The comparison gives the launch a numerical reference point, but it remains Pienomial’s presentation of the results—not an independently established conclusion that AT0M will outperform Jev on a business’s own work.

Pienomial also reports reproducible responses and a response time under 16 milliseconds in a hosted test using an NVIDIA T4 graphics processing unit. That timing belongs to the named test setup. It should not be read as a promised response time across every Intel, Apple or NVIDIA configuration the company says it supports.

For a prospective customer, the unresolved question is how those test results translate to its particular decisions. The benchmark measures agreement with reference answers; the latency figure measures response time on one hosted setup. Neither figure, by itself, settles whether AT0M is accurate enough for a given workflow or which cases should still require human review.