Skip to content
fourthSol

Search the higherMark suite

Jump to any page

on-device AI

The model

In development

A domain-tuned model that runs on your machine — offline by default, Lightning-native, with a cloud boost only when you ask for one. Not an API key pointed at someone else's weights.

Mainstream page

The short version of this page, for everyone else, lives on the higherMark domain.

highermark.xyz/model

At a glance

Class
7–12B, domain-tuned
Base
Permissively licensed open weights
Method
LoRA / QLoRA fine-tune
Packaging
Quantised — Q4/Q3 GGUF
Targets
Apple Silicon · iOS
Default
Local inference, offline

Throughput

Local is fast enough. The cloud is faster.

A well-tuned 7–12B model clears 50+ tokens per second on a modern Mac and stays usable offline on a phone. Cerebras-class speed stays where it lives — in the cloud, on specialised hardware — which is exactly why it is optional here.

  • CerebrasWafer-scale, cloud only — the optional boost~3,000 tok/s
  • Cloud averageTypical hosted inference across providers50–200 tok/s
  • Mac · localQuantised 7–12B on modern Apple Silicon50–80 tok/s
  • Phone · localLighter weights — slower, but usable with no network10–25 tok/s

Local figures are for quantised weights on Apple Silicon; phones run lighter models and land lower. Reference numbers, not guarantees — we publish measured throughput per build.

How it's built

Tuned, not rented

Five steps from an open base to weights we own.

  1. Choose the base

    An open model whose licence permits commercial derivatives. That single choice decides both the ceiling on performance and whether the result is actually ours.

  2. Build the dataset

    The real work. Domain data collected and cleaned into instruction–response pairs, augmented synthetically only where the licence allows, versioned and kept private. Quality over quantity.

  3. Fine-tune

    LoRA / QLoRA on rented GPUs, validated against held-out domain tasks and iterated. The resulting adapter is a derivative work we own, subject to the base licence.

  4. Quantise and integrate

    Phone-friendly and macOS-optimised builds embedded in the clients, so inference is local by default. Throughput measured on real target devices rather than claimed.

  5. Deploy and iterate

    Ship inside the alpha, gather privacy-respecting usage signals, retrain periodically on new proprietary data. Fully offline and hybrid modes both stay supported.

Ownership

What is actually ours

The honest version, because the answer depends entirely on the base licence.

The fine-tuned weights, the training code and the dataset are company property. The base model's licence still governs attribution and usage restrictions, and we comply with it — that is why the licence is the first decision, not the last.

Training data stays private and everything is documented. Stronger IP claims would mean continued pre-training or a smaller model from scratch, both expensive; that is a later question, and one for counsel once the first working derivative exists.

Privacy

Nothing leaves unless you send it

Offline-first is the default, and the network is the exception.

Fully offline

The model runs on your device, on your data, with no network at all. Ask questions with no bars and no wifi — the same way you can pay with none.

Hybrid

Local by default, with a Cerebras boost for the heavy asks. You choose per request, and you can see which one answered.

Usage signals, where collected at all, are privacy-respecting and opt-in.

Elsewhere

The mainstream pages

Cleaner, shorter versions of this and the desktop app, on the higherMark domain.

Next