Skip to content

§ 05 · Practice

Frontier-quality answers, laptop-sized models.

For narrow, well-defined tasks, a small model trained on the right data beats a large general one, and it runs on the MacBook already on your desk. We select, fine-tune, quantise and package models that work fully offline.

§ 05.1 / Scope

Capabilities.

01

Model selection

Baseline evaluations across open model families (Qwen, Llama, Gemma, Phi and others) on your actual task, with licensing checked.

02

Fine-tuning

Supervised and preference tuning with LoRA and full fine-tunes; distillation from frontier models when you lack labelled data.

03

Quantisation & packaging

4-bit and 8-bit quantisation for MLX, llama.cpp and Core ML. Accuracy loss measured, not assumed.

04

Native deployment

Signed macOS apps, command-line tools or a local API server. Works on a plane, in a clean room, or behind an air gap.

05

Profiling

Tokens per second, memory footprint, thermal and battery behaviour on the M-series chips your team actually uses.

06

Lifecycle

An evaluation and update pipeline so the model improves as your data does, without regressions.

§ 05.2 / Output

Deliverables.

  • Model weights, the training recipe and an evaluation report
  • A signed macOS app or CLI, or a local server
  • Fine-tuning and evaluation code you own
  • An update playbook for the next version

Tools & Standards.

nf-core and community standards where they exist, custom software where they do not.

  • MLX
  • llama.cpp
  • Core ML
  • PyTorch
  • Hugging Face
  • LoRA / QLoRA
  • Unsloth
  • Swift
  • lm-eval

§ 05.3 / Questions

Frequently asked questions.

Why a 3B model instead of a frontier API?
Privacy, cost and latency. A 3B model in 4-bit fits in about 2 GB of memory and runs at tens of tokens per second on an M-series chip, with no per-token cost and no data leaving the device. For narrow tasks, fine-tuned small models match or beat general large ones.
What tasks are a good fit?
Classification and extraction, summarising reports, protocol Q&A, variant and literature triage, code and command generation for your tools. Open-ended research questions still belong with a frontier model.
Which Macs are supported?
Any Apple Silicon Mac. We target 8 GB machines as the floor and profile on M1 through M4 families.

§ Practice Scoping · 05

Scope a project in on-device ai models.

Tell us about your assay types, dataset scale, and timelines. We reply with a preliminary scoping memo within two working days.

Scope On-device AI Models →