Releases: johnbean393/KeyType
Releases · johnbean393/KeyType
1.4
Update
New Features
- Manual exclude app
- Settings window resize
- Latency snapshot comparison
- 1.11x speedup via zero-copy prefix checks
- 1.06x speedup via numeric mid-word suppression
- 1.18x speedup via adaptive beam width
- 1.06x speedup via guarded mid-word decode budget
- 2.39x speedup via default mid-line suppression
- 1.33x speedup via retokenized anchor history
- 1.27x speedup via overlapped debounce generation
- 1.18x speedup via tighter adaptive debounce
- 1.1x speedup via batched FIM rerank
- Improved prompt logging
- KeyTypeBench results
- Eval harness
Bug Fixes
- Slack completion insert
Requirements
- macOS Sonoma (14.0) or later
- Apple Silicon Mac recommended for on-device model inference
1.3
Update
New Features
- End to end latency profiling
- Improve non-Latin suggestions
Bug Fixes
- Notion compatibility
- Multiline text field wrapping
- Obsidian compatibility
- Slack and Discord compatibility
- Invasive accessibility interference
- Improve Chromium app focus capture
- Dynamic web app detection
Requirements
- macOS Sonoma (14.0) or later
- Apple Silicon Mac recommended for on-device model inference
1.2
Update
New Features
- Branch promotion rollback support
- Branch promotion cache
Bug Fixes
- Screencapture compatibility
- Default OCR to off
- ACPF profile build for duplicate-byte tokenizers (Gemma) (#9)
Requirements
- macOS Sonoma (14.0) or later
- Apple Silicon Mac recommended for on-device model inference
1.1
Update
New Features
- Show app icon for windows
- FIM quality improvements
- Beam search early exit
Bug Fixes
- Mid-word healing extra space
- Type-through race condition
- Native app ghost text redrawing
- ACPF profile build error logs
Requirements
- macOS Sonoma (14.0) or later
- Apple Silicon Mac recommended for on-device model inference
1.0
KeyType
On-device, system-wide tab autocomplete for macOS.
Features
- System-wide ghost-text completions accepted with Tab
- Fully on-device prediction with a local LLM (no network calls)
- Short, cursor-anchored continuations that prefer silence over a wrong guess
- App-aware insertion and overlay behavior
- Private by default: clipboard, screen/OCR, and writing-history context are opt-in
Requirements
- macOS Sonoma (14.0) or later
- Apple Silicon Mac recommended for on-device model inference