Choosing a model

Every model in the catalog — plus the Apple Intelligence system model — is measured on the same held-out eval: first-word accuracy on real human text, where staying silent counts as a miss.

Dumbbell chart: on English all nine models score 24-28% first-word accuracy, a tie; averaged over 16 other languages they fan out from 22.6% (Gemma 4 E4B) down to 8.5% (Apple Intelligence)

The default used to be chosen from your keyboard layouts; since 2026-07-25 it is sized to your Mac instead — the strongest tier whose resident model stays within about a quarter of physical memory:

Mac memoryDefaultResident (recommended config)
8 GBQwen3.5 2B≈1.6 GB
16–18 GBGemma E2B 4-bit≈3.5 GB
24 GBGemma E2B 8-bit≈5.0 GB
32 GB and upGemma E4B 6-bit≈6.8 GB

Language dropped out of the rule because no per-language gap between the small models survives the significance bar (MiniCPM5’s English lead over Qwen3.5 is a trend at p=0.014, Russian a tie at p=0.79, Japanese p=0.064), while the chord measurements split hard against the old English/Russian default: MiniCPM5’s fix is inert at 4%. It remains the manual pick for the lowest latency (49 ms) and the best reply drafts (56%); E4B 8-bit remains the manual best-quality pick (it ties 6-bit at p=0.052 for 1.8 GB more). Your keyboard languages still drive the persona and the per-language menu nudge, and everything else is one click in Settings → Model, where each entry ships its own eval-backed recommended settings and the catalog re-ranks for any of the 19 evaluated languages.

Model tab: an interactive speed by accuracy map of the catalog with eval-backed presets and a ranked list

The same data live in the app — a speed × accuracy map with one-click presets and a per-language accuracy axis.

Chart method: equal-weight mean over cs de es fr it ja ko nl pl pt ro ru sv tr uk zh, matched register cells only (Arabic and Hebrew were measured later, on their own set, and are not in this chart). Absolute values are not comparable between languages (zh/ja are character-masked), so the mean measures spread across languages, not skill at any one of them. Every language cell is 280 held-out rows — English 560, Russian 689 — which is ±5 pp near 20% and ±3 pp at Russian’s size; a gap smaller than that is not a result. The app shows the same tolerance and sample size beside each figure, and lets you pick any single language as the axis rather than reading an average.

Fixing and replying

The two chords — fix the selection (⌥Tab) and draft a reply from the screen (⌥⇧Tab) — don’t run on the model you picked: they run on its instruct sibling, so the numbers that matter are the sibling’s. Measured 2026-07-25 on their own held-out sets: 510 single-typo sentences plus 170 clean controls across 17 languages for the fix (through the full app pipeline, gates included), 570 on-screen conversations across 19 languages for the reply.

Runs the chordsLoaded by (recommended settings)Exact fixTouches clean textFix p50Usable repliesReply p50
Gemma E4B-it 4-bitGemma E2B 8-bit; floor for both E4B tiers¹65% [60–69]21%1.5 s33% [30–37]0.5 s
Gemma E2B-it 4-bitGemma E2B 4-bit50% [45–54]29%0.4 s32% [28–36]0.4 s
Qwen3.5 2B (itself)Qwen3.5 2B — the 8 GB default32% [28–37]28%0.3 s46% [42–50]0.4 s
Bonsai 4B (itself)Ternary Bonsai 4B21% [18–25]14%0.3 s21% [18–25]0.3 s
Qwen2.5 0.5B-itQwen2.5 0.5B11% [9–14]19%0.2 s38% [35–42]0.3 s
MiniCPM5 1B-itMiniCPM5 1B4% [2–6]14%0.7 s56% [52–61]1.8 s

Three things worth knowing before reading it as a ranking:

¹ The E4B 8-bit and 6-bit tiers run the 6-bit instruct sibling, which the sweep didn’t cover — the 4-bit row is its measured floor. Method: totals over 510/570 rows carry ±4 pp; per-language cells are 30 rows (±17 pp) and are only used here for direction, never for claims. Fix numbers are from the 2026-07-25 rerun after the run-on trim and turn-marker stop landed — the first pass lost 39% of the E4B sibling’s rows to same-line junk, which is what the rerun recovered. Sets: eval-correct (one real typo per sentence, Tatoeba register), eval-reply (screen-shaped conversations, right-language/no-echo contract).

What a model costs you

Disk and memory track each other: the default holds ≈1.6–6.8 GB resident — sized to about a quarter of the Mac’s memory, per the ladder above — the catalog goes down to ≈1 GB, and Gemma 4 E4B 8-bit tops it at ≈8.6 GB. Warm completions run 49–145 ms across the local models, against 430 ms for Apple Intelligence, which downloads nothing and holds no app memory at all.

That resident figure is what a model holds while you type, not all day: it unloads itself after five idle minutes and reloads on your next keystroke, and unloads early if macOS reports memory pressure. Using ⌥Tab rewrites can add a one-time instruct sibling on some models — ≈2.2 GB on MiniCPM5, up to ≈5 GB on a Gemma running Base style; none on Qwen3.5 or Bonsai, which correct with themselves, and none on a Gemma tier in its recommended Instruct style, where the loaded model already is the instruct build.

Quantization tiers, the engine split and the confidence gate are in Architecture. Training a model on your own writing is in Fine-tuning.