TabType v2.0.0 replaces the completion engine from the ground up. v1 asked a chat model to "continue the user's text", and too often it replied, rambled, or guessed. v2 runs a base language model through llama.cpp: it simply continues your writing, and a new decoder only shows a suggestion when the model is actually confident.
📈 Measured, not guessed
On TabType's offline eval set (192 real-world typing cases, tabtype-eval):
| v1 | v2 (Qwen3-4B base) | |
|---|---|---|
| Next word correct | 37% | 56–58% |
| Precision (shown suggestions that were right) | 38% | ~70% |
| Wrong suggestions shown | 61% | ~22% |
| Median latency | ~300 ms | ~80–100 ms |
🧠 New engine
- llama.cpp + GGUF base models. Qwen3-4B on 16 GB+ Macs, Qwen3-1.7B on 8 GB. Gemma 4 and smaller Qwen models are selectable in Settings → Model.
- Confidence-gated decoding. Every suggestion carries the model's own probability. Weak guesses are never shown, and multi-word phrases only extend while confidence holds.
- Token healing. Mid-word suggestions finish the word you're typing instead of starting a new one.
- Prefix-cached prompts. Context is laid out so the model reuses its cache across keystrokes, which keeps it fast.
✍️ Looks native
- Font-fitted ghost text. TabType renders candidate fonts and matches them against the field's pixels, so the ghost uses the app's own font, size, and baseline.
- Tab-accept keeps the rest of a suggestion in place instead of recomputing it.
🙋 Learns from you (opt-in)
- Learn from your writing. An encrypted, on-device index of what you've written nudges suggestions toward your names, phrases, and sign-offs. It's off by default and can be turned on per app. Erase it any time in Settings.
- Voice adapter. Advanced users can load their own LoRA adapter (GGUF) to steer style.
🧹 Removed
- MLX and the Apple Intelligence engine are gone. Base models complete text measurably better, and the app no longer has any third-party Swift dependencies.
- Temperature / max-token settings were removed; the decoder decides length from confidence.
- Phrase memory and the old typing history are replaced by Learn from your writing, which imports your existing history once.
Install / Update
Download TabType-2.0.0.dmg below and drag it to Applications, replacing the old copy. Your permissions are preserved.
Updating from 0.1.x: v2 uses a new model format, so it downloads a new model on first launch (~1.1–2.5 GB, depending on your Mac's RAM). To reclaim space, delete the old mlx-community… folders inside ~/Library/Application Support/TabType/Models (keep the .gguf files: that's v2's model) and any models--mlx-community… folders in ~/.cache/huggingface/hub.
New install? macOS will warn about the unnotarized app: System Settings → Privacy & Security → "Open Anyway", then grant Accessibility. Full steps are in the README.
Apple Silicon Mac, macOS 14+.
SHA-256 9509a310ce630535fccf24b5341a454c3fb7760b58afc5440867b7aa686d7bb5
⚠️ Still an alpha. If ghost text sits wrong or suggestions miss in a particular app, bug reports are gold.