Tech overview

What Neutron Studio is

Neutron Studio is a local AI workstation in private beta for Apple Silicon Mac. It remains part of NeutronTech’s research program, not a second flagship. The current candidate brings conversations, documents, spreadsheets, and source-referenced answers into a native workspace.

Privacy is the default, not a setting you have to find.

Jimmy and the foundation

Jimmy is the agent in Neutron Studio. Neutron Engine provides the underlying conversation, document, retrieval, context, and task-execution capabilities. The same foundation is intended for adoption by Hubyn, with Hubyn retaining its own identity, permissions, and workspace rules.

• The engine owns durable records, source snapshots, indexes, context construction, and task execution.
• Neutron Studio owns Jimmy’s behavior, interface, permissions, and product defaults.
• Planned adoption by another product is not evidence that integration has already happened.

Documents and checked answers

The current candidate preserves imported originals and conversation records locally. Its document workflow includes source references, complete-document review, scanned-PDF text extraction, spreadsheet worksheets, and checked calculations. Coverage and extraction limits matter: a source reference does not guarantee that an answer is correct.

• Format-aware processing extracts document text and spreadsheet data before selected evidence enters an answer.
• Coverage, omissions, and failures must remain visible rather than being presented as a complete review.
• Checked calculations use data operations; model interpretation can still be wrong.

Local model and Mac requirements

The next v1.0 candidate is being prepared around one bundled model, Gemma 4 E2B through MLX, targeting Apple Silicon Macs running macOS 26 or later. Model replacement, hardware qualification, and native testing must finish before this is described as a released configuration. Additional models and expanded model selection are planned.

The model’s working memory isn’t a fixed number. It’s set from the model and from your Mac’s memory, so the same app does more on a bigger Mac with nothing to configure.

• The intended configuration does not require Apple Intelligence.
• Artifact identity, redistribution rights, loading, memory use, streaming, and cancellation require qualification.
• Results measured with the previous model do not qualify its replacement.

Conversations with a record

The current candidate stores conversations and sources locally so they survive restart. Conversation management includes branching, retrying, archiving, and Markdown export. Original messages and imported source bytes remain separate from replaceable indexes and summaries.

Context is sized to your Mac, not fixed. Neutron Engine reads how much the model can hold and how much your Mac’s memory allows, and uses that. A Mac with more memory holds more of the conversation at once. Long conversations are condensed as they grow, and nothing you said is dropped: your corrections and open tasks are always kept, and earlier details are retrieved when they matter.

Permissions and bounded tasks

The engine is designed around explicit source scopes, host-granted capabilities, and resource budgets. A model does not grant itself access or establish that an action succeeded.

• Only authorized sources should enter a task.
• Work must expose progress, cancellation, results, and failures.
• Reviewable editing and document creation remain planned; current candidate capabilities do not establish those workflows.

Local data work, with honest boundaries

The current candidate includes spreadsheet worksheets and checked calculations over supported data. Deeper extraction and advanced comparisons remain development goals.

• A checked calculation does not make this an Excel formula engine or guarantee Office round-trip fidelity.
• Supported formats, file sizes, extraction coverage, and answer quality require evidence from the exact candidate.

Permissioned online work

Local work does not require a cloud AI service. Optional web research is planned and would require permission to send a search request online. It is not part of the current qualified offering.

• An approved online search can disclose information included in its query.
• Planned search must provide attributable, dated evidence and remain separate from local-only work.

Your data lives on your Mac

The local core keeps conversations and source files on the Mac. Network activity for intended updates or future optional search is not the same as sending a local workspace to a remote AI service.

• Deletion behavior must be tested across original records and derived data.
• Do not equate hiding or archiving a conversation with erasing it.

The Mac can join. It can never quietly become the owner.

From answers to finished work

Planned next steps include deeper documents, reviewable editing and creation, local-first team workspaces, and later audio and visual processing.

• Team sharing, optional sync, and distributed execution are future scope, not current privacy or availability claims.
• Each milestone requires observable workflows and acceptance evidence before being described as delivered.

Checking our work

The private-beta candidate is undergoing testing and distribution qualification. Direct distribution and signed in-app updates are the intended release path.

• A released configuration requires native QA, model requalification, verified signing and notarization, and clean-Mac installation and update checks.
• Do not describe the candidate as released, notarized, or available through Homebrew until those steps are verified.
• Private vulnerability reports should identify the affected candidate and reproduction steps without exposing private documents or signing material.