Here is **LLMVoice**, a technical demo designed to evaluate the new iOS 26 speech recognition APIs (**SpeechAnalyzer / SpeechTranscriber**) and the execution of **open-weight LLM models** (quantized in **4 bits**) locally on an iPhone that supports the **MLX** framework. You can either summarize a transcription or send a prompt directly to the model. Let’s be honest — models this compressed aren’t particularly useful, but it’s still fun to watch them hallucinate. That said, they’re powerful enough to bring my old iPhone 13 to its knees! Here are the available downloadable models:
**Gemma3_1b** (mlx-community/gemma-3-1b-it-4bit)
**Qwen25_05b** (lmstudio-community/Qwen2.5-0.5B-Instruct-MLX-4bit)
**Llama32_1b** (mlx-community/Llama-3.2-1B-Instruct-4bit)
## Source Code ## System Requirements iOS 26.0+ / iPadOS 26.0+ ### For Apple Intelligence iPhone 15 Pro or later Apple Intelligence enabled in Settings ### For local MLX models iPhone 12 or later (A14 Bionic chip or higher).
iPad Pro 2021 or later (M1 chip or higher).
iPad Air 5th generation or later (M1 chip or higher).
Metal GPU with specific features: air.simd_sum, rmsfloat16 kernel support.
### Storage App size: about 50 MB Model downloads (MLX only): Qwen2.5 (0.5B): about 150 MB
Gemma 3 (1B): about 300 MB
Llama 3.2 (1B): about 500 MB

--- ## 🛟 Support Need help or want to share feedback? Email [support@andrefrelicot.dev](mailto:support@andrefrelicot.dev).