Jexxa Ships Local Dictation for Mac With Offline Inference and Flat Pricing

Jexxa is Mac dictation software that runs transcription on-device with the model stored locally on the user's disk. The product page lists the current release as version 0.1.10 as of September 15, 2026 jexxa.org.

The system requirements are narrow. Jexxa requires macOS 14 or later and Apple silicon. No Intel build is listed. That constraint simplifies the inference target to Neural Engine and unified memory architectures, and it avoids the driver and performance variance of supporting older Macs.

Privacy and connectivity are handled as a local execution problem. The website states that no audio or transcript is uploaded from the Mac, and that the software works offline without Wi-Fi. For tech-literate users, the implication is straightforward. Dictation input never leaves the endpoint for server-side decoding, which removes network round-trip time from the latency budget and keeps sensitive speech out of transit and remote logs.

Latency is specified in user-perceived terms. Jexxa's website states that most dictations are typed within a fifth of a second of releasing the dictation key. The software also provides a live preview of transcription while the user speaks. The combination matters for correction workflow. Streaming partial hypotheses lets the speaker catch errors early, while fast commit on key release keeps dictation usable inline in editors, terminals, browsers, and other text fields without breaking typing flow.

Editing is voice-driven as well as keyboard-driven. Jexxa supports the commands "JX minus one," "JX minus two," and "JX clear" to delete the last line, last two lines, or all dictation. Fixed-phrase commands of this kind are easier to recognize reliably than open-ended edit instructions, and they give the user a deterministic way to roll back misrecognitions without reaching for selection keys.

Personalization is explicit and reversible. Jexxa learns corrected names and jargon after a user fix and allows the learning to be undone. That design addresses a familiar failure mode in dictation systems. Proper nouns, product names, API identifiers, and domain terms are long-tail vocabulary that a base model will miss. Learning from a correction closes that gap for repeat use, while an undo path provides recourse when a correction introduces a persistent error.

Distribution reflects memory constraints on shipping local models. Jexxa is offered in two variants: full JEXXA with a 3.1 GB download for Macs with 16 GB or more memory, and JEXXA Small with a 2.0 GB download for 8 GB Macs. The split lets users trade footprint against available unified memory, which is a practical concern when the transcription model shares RAM with IDEs, browsers, containers, and other resident workloads.

Pricing is flat. Jexxa costs $8 per month plus tax with unlimited dictation and no per-minute billing.

Looking at what this means for deployment on laptops, the interesting trade is local resource consumption in exchange for determinism. Cloud dictation offloads memory and compute but couples everyday input to connectivity, queueing, and data-handling policy. A 2.0 to 3.1 GB local package reverses that. The user pays once in disk and working memory, then gets consistent offline behavior and sub-second commit. In my view, that is a reasonable exchange for professionals who dictate in source code discussions, customer notes, medical or legal drafting, or any setting where transcripts are sensitive and Wi-Fi is not guaranteed.

Worth flagging for evaluation is how the learning layer behaves over time. Short-term adaptation to names and jargon is useful. Long-term value depends on whether learned terms remain scoped correctly, survive updates, and stay easy to inspect and remove. Jexxa's inclusion of an undo control suggests attention to that problem, and hands-on testing with dense technical vocabulary would show how well it holds up.