https://appagg.com/developer/acmeup-inc/points/?hl=ko ko AppAgg.com Sun, 04 Jan 2026 15:26:11 GMT Sun, 04 Jan 2026 15:26:11 GMT https://appagg.com/developer/acmeup-inc/points/?hl=ko https://appagg.com/i/apple-touch-icon.png 동작: +1 포인트  
사용자를: cahiba7911@agenra.com  
등록 날짜: 2026-01-04 15:26:11 UTC  

가격: 무료   
목록: 0 + 0   
포인트: 1 + 10 (4.6)   
순위: 0   
댓글: 0   
버전: 2.0.3  
게시자: AcmeUp Inc.   
업데이트: 2026-04-09 09:08:01 UTC  
출시일: 2025-07-18 07:00:00 UTC  
카테고리: iOS 앱 › 생산성  

기술: 
NEW: Gemma 4 and Bonsai - Next-Gen Local Models Gemma 4 (E2B, E4B) runs on-device with vision (Gemma 4V) and full tool-calling. Bonsai-8B uses PrismML 1-bit quantization (Q1_0), ... 

새로운 기능: 
FEATURES:
- Device RAM Management: New side menu page shows live RAM stats, loaded model list by type, memory tips, and an "Unload All" button with haptic and toast feedback.
- Real-Time Streaming ASR: FluidAudio and MLX Audio engines now show a live transcript panel with partial hypotheses during recording; non-streaming engines (WhisperKit) are unchanged.
- Gemma 4 Support: Local inference supports Gemma 4 and Gemma 4V (vision) with full tool-calling on E2B/E4B models.
- Bonsai-8B 1-Bit Quantization: PrismML Q1_0 1-bit quantized models run on-device with full Metal acceleration.

IMPROVEMENTS:
- Navigation bar titles added to MLX Audio and FluidAudio model list views.
- EXIF metadata (camera, date, lens, GPS) from attached photos is now sent to the AI.
- iPad chat query editor stays open while scrolling; keyboard no longer dismisses on scroll.
- Tool sheet shows a "Manual | Auto" toggle with an explanation card in Auto mode.
- Small context models (4K-8K) get compact tool lists; large context models (64K+) get richer descriptions.
- Local model default context scales with file size: under 1GB = 20k tokens, 1-2GB = 14k, larger = 8k.
- Swipe-to-open-menu gesture active in list views, disabled in settings with sliders.
- FluidAudio upgraded to tag-20260406: Japanese and Mandarin ASR models, PocketTTS session prefill, crash fixes.
- MLX Audio Swift upgraded to tag-20260403 with MLXFast for Metal GPU acceleration across all STT models.
- Embedder no longer loads on every chat start; loads only when past messages or documents exist.
- MLX model loading is 10-15x faster with a streamlined tokenizer.
- llama.cpp b8690: Q10 quantization, HunyuanOCR, Granite 4.0 chat template, Gemma 4 and KV cache fixes.

BUG FIXES:
- Voice picker now shown for all TTS models (FluidAudio and MLX Audio) that offer multiple voices.
- ASR falls back to the per-engine language setting when no language is explicitly provided.
- ASR test view now shows error messages in red instead of dismissing silently.
- "Set as Default ASR" now shows a confirmation alert and updates state immediately after confirmation.
- Switching ASR engines now correctly clears the previous model's default flag.
- Language picker added to the ASR test widget embedded in model detail previews.
- Spoken numbers, currencies, and percentages now transcribed as formatted text ("$200", "50%", "3rd").
- MLX Audio TTS models now store voice preferences independently per model.
- TTS and ASR re-verified after storage clear; app falls back to system engine when files are missing.
- Stopped generation no longer leaks tokens into the next query; in-flight network requests cancelled on stop.
- MLX vision models no longer crash with repetition, presence, or frequency penalty settings.
- Council chat history now loads correctly after app restart.
- Writer chapter text selection no longer crashes when the copy/paste menu appears.
- iPad Council Chat input bar now rises above the keyboard when tapped. 

AppAgg: 
[ Privacy AI: Powerful chatbot ]

❤️ Support AppAgg →
]]> Sun, 04 Jan 2026 15:26:11 GMT