Skip to content

Research · November 8, 2026

Which local model fits a 16 GB Mac

We compared open models by memory tier for extraction and tool calling. The scores come from vendor model cards, so we call our defaults provisional until we test Turkish and English ourselves.

A local model keeps text on the computer, but only if it runs well on the computer you have. Choosing one is a memory question first and a quality question second. This note records how we compared candidates and how much of that comparison we trust.

The task we chose for

The research note states the task profile: extract tasks, decisions, dates, risks and amounts as JSON from long chunks, write summaries, call tools, and work in Turkish and English. That is narrower than "chat", and it is why we looked at tool-calling scores and not only general knowledge scores.

What the model cards report

The note read the model cards directly on the day of research, and marked each figure by how it was obtained. For two sizes of the same family, the cards report a function-calling score (BFCL-V4) of 66.1 for the nine billion parameter model against 50.3 for the four billion model. All five models in the table carry an Apache licence on their cards.

Those are the vendor's own numbers, measured with the vendor's harness. The note also flags that older scores from other benchmark versions are far lower and must not be compared with them.

Memory tiers

Files at four-bit quantisation take about 0.6 GB per billion parameters plus working memory, which is the note's own arithmetic. From that, the note proposed these defaults:

  • 16 GB Mac: the nine billion parameter model, with the four billion model as the fallback.
  • Windows without a discrete graphics card: the four billion model, because the note's source data puts the larger one at roughly eight to ten tokens a second on a processor alone.
  • 32 GB Mac: a larger mixture-of-experts model, with only a few billion parameters active per token.

The shipped app works from a catalog, and the catalog recommends Qwen3 8B on a 16 GB Mac; the research proposal is a different model family and is not yet the default (docs/methods/site/local-model-sizes.md). The app measures your computer and shows, per model, whether it fits. We keep the two statements apart on purpose.

What a local model does and needs

A local model runs on this computer and sends no text over the internet. It runs through Ollama, which must be installed and running on the same Mac. See the local-first entry and how to keep business data on your computer.

What we could not establish

The note lists its weak spots first, and we repeat them.

  • No Turkish-specific benchmark was found for any model, so Turkish quality is indirect evidence from multilingual averages.
  • Speed figures for Apple Silicon come from a benchmark of an older model; the note scaled them by its own estimate.
  • Several candidates had no function-calling score at all, so they are listed as not evaluated, not as worse.
  • Windows data without a graphics card is thin.

What we decided

The defaults stay provisional. The note's next step is a bake-off on Turkish and English chunks with a JSON grammar, scoring field-level accuracy, JSON validity and tool-call parsing, plus real speed on a few real machines. Until it runs, we do not publish a ranking as a finding, and the app decides from your own computer's measurement.

Where this comes from

  • docs/research/local-llm/LOCAL_LLM_TIERS_2026-10-30.md, read directly from model cards and discussions on 2026-10-30; speeds and sizes marked as estimates in the file.
  • docs/methods/site/local-model-sizes.md, dated 2026-11-06, for the app's catalog.

Sources