LLMVoice vs LLM Bench

LLMVoice — iOS 26 Speech & On-Device AI Showcase A technical showcase of Apple's latest iOS 26 technologies for speech recognition and on-device AI, including MLX Swift for running open-source language models locally on Apple Silicon. Built for Developers & Early Adopters LLMVoice is made for developers and AI enthusiasts who want to explore Apple's newest frameworks and on-device AI capabilities, end to end, on a real device. What It Does • Real-time voice transcription using iOS 26 speech APIs • AI-powered summarization with multiple model options • Generated code shown with syntax highlighting • Live HTML preview for generated content, with built-in diagnostics • Optional "thinking" view: see the model's reasoning, or hide it for a clean output • Modern SwiftUI interface with smooth streaming updates • 100% on-device processing for complete privacy Key Technologies MLX Swift Framework Apple's MLX (Machine Learning Extensions) enables efficient inference of quantized language models on Apple Silicon with optimized GPU acceleration. Apple's On-Device Foundation Models On supported devices (iPhone 15 Pro or later), use Apple's built-in system models for summarization — no extra downloads required. Open-Weights Models via MLX Choose from multiple open-source models running locally: • Qwen2.5 (0.5B) — 150MB, 29+ languages, 32k context • Gemma 3 (1B) — 300MB, 140+ languages, 32k context • Llama 3.2 (1B) — 500MB, 8 languages (EN/ES/FR/DE/IT/PT/HI/TH) All models use 4-bit quantization for mobile efficiency. Token limits are tuned per model for reliable results. SpeechAnalyzer & SpeechTranscriber (iOS 26) Low-latency, real-time on-device speech recognition using Apple's newest APIs. Modern Swift & SwiftUI Built with Swift 6, the Observable macro, async/await, and actor-based concurrency. How It Works 1. Select an AI model (Apple's on-device models, or download an MLX model) 2. Tap the microphone to start recording 3. Speak naturally — real-time transcription appears 4. Tap stop when finished 5. The model generates a concise summary instantly 6. Summaries are saved locally and persist between sessions 7. Swipe to delete, tap to view the original transcription Privacy-First Everything runs on-device: • Audio never leaves your device • Local speech recognition and AI inference • No cloud processing or internet required (except model downloads) • No tracking, no data collection Requirements • iOS 26.0+ (iPadOS 26.0+ for iPad) • Apple Silicon recommended • Storage for models (150MB–500MB each) • Microphone and speech recognition permissions Good to Know • Models are downloaded on first use from Hugging Face • Apple's on-device models are available immediately on compatible devices, without downloads • Primarily optimized for English; other languages vary by model Perfect For • Developers exploring iOS 26 and MLX Swift • AI enthusiasts interested in on-device LLMs • Students learning SwiftUI and modern concurrency • Anyone curious about running open-weights models on mobile Under the Hood Demonstrates MLX Swift integration with Apple frameworks, including model downloading, GPU memory management, streaming inference with real-time metrics, and efficient on-device transformer execution. Mac Catalyst Compatible Runs on macOS 26+ with Apple Silicon for a unified cross-platform experience.

LLM Bench is an iOS test bench for developers comparing on-device LLM behavior under real device constraints. Run the same prompt against selected local model targets and compare the evidence before choosing what belongs in your app. LLM Bench focuses on practical signals: load time, first output time, generation time, output speed, memory, thermal state, model readiness, raw output text, failures, and saved run history. Use LLM Bench to: - Compare GPU-capable local model runs on iPhone and iPad. - Run repeatable prompts against selected model targets. - Inspect model output, timing, memory, and thermal state in one place. - Save benchmark runs for later review. - Export benchmark evidence as CSV. - Add models from Hugging Face URLs, direct file URLs, or local imports when the artifact format is supported. LLM Bench is built for technical builders who want a compact measurement tool, not a leaderboard or generic AI chat app. It helps you answer the practical question: which local model behaves well enough on this device for this use case? The free tier includes one completed benchmark run. Subscribing unlocks ongoing benchmarking, saved comparison work, and export workflows. Premium is available as an auto-renewing subscription (monthly or yearly), or a one-time Lifetime purchase. Payment is charged to your Apple ID at purchase; subscriptions renew automatically unless canceled at least 24 hours before the period ends. Manage or cancel anytime in your App Store account settings. Privacy Policy: https://thedandano.github.io/michilotl-studios-docs/apps/llm-bench/privacy/ Terms of Use (EULA): https://www.apple.com/legal/internet-services/itunes/dev/stdeula/