AI Briefing#on-device-ai
The latest AI and developer news about #on-device-ai, with the original source and a short summary.
The latest AI and developer news about #on-device-ai, with the original source and a short summary.
The model achieves 80+ tokens/s decoding and 29GB peak memory usage on compatible laptops, outperforming several cloud models on coding and reasoning benchmarks.
Tether AI released BotPsy-0.8B, based on Qwen3.5-0.8B, demonstrating superior tool-use performance compared to 2B models.
The 74M parameter codec uses Grouped FSQ to enable parallel prediction, outperforming Mimi and NanoCodec in reconstruction quality and speech generation metrics.
The open-source model achieves SOTA tool-use performance for its size, scoring 45.0 on average compared to 38.8 for Qwen3.5-2B-Think.
The new backend allows a single .pte file to run on Chrome, Safari, and other WebGPU browsers without conversion to other formats.
The 131M parameter model supports 6 languages and 14 voices, achieving 30ms warm start on M5 Max and 0.85% WER on the Open TTS Leaderboard.
Gradium's Phonon TTS, available in 31m and 40m sizes, tops the Hugging Face Open TTS Leaderboard's CPU streaming benchmark for time-to-first audio.
Software Mansion's open source app Private Mind now supports LFM3 models, achieving ~40 tok/s on iPhone 16 Pro via a custom ExecuTorch export.
The ~70 MB model uses 200+ trained heads on a shared encoder to output transcripts, linked names, and metadata every 100 ms.
The 1B OCR model runs locally on mobile devices while the 2B verifier model checks output accuracy against original images.
The M6-based iMac is up to 5 times faster in on-device AI first token generation compared to the M4.
Apple has launched Korean language support for Siri AI through the 27.2 update, enhancing privacy through on-device processing.
Apple's 20B 'Core Advanced' model runs on macOS 27 using approximately 2.0–3.4 GiB of RAM by storing weights on flash, consuming about one-third of the energy of comparable open models.
The app uses GUI-Owl 1.5 2B to operate Android apps locally, achieving 7.3s step times on M2 CPU and 236ms vision encoding on Galaxy S25 NPU.
The model supports a 1,048,576-token context and offers controllable thinking budgets to balance reasoning quality and latency.
The new model retains the Spark-X2.5-4B architecture while delivering state-of-the-art performance among similarly sized open-source models.
Cristiano Amon stated at Snapdragon Summit 2026 that leading AI companies require the ability to run 100-billion-parameter models on devices continuously by 2028.
The framework improves reasoning accuracy in 1.5B parameter models by integrating In-Context Learning into Test-Time Reinforcement Learning.
The new 35B-A3B MoE model utilizes Lookahead Expert Scheduling to enable local vision-language processing on smartphones.
The open-source app uses a hidden WebView and local embeddings to retrieve web content without any backend or search API.
The open-weight model achieves 184 tokens/s on an Apple M5 Max and scores 48.09 on LiveClaw, outperforming larger models in agentic tasks.
The tool transcribes one hour of audio in approximately 13 seconds on a base M3 MacBook Air using the Neural Engine.
The new open-weight model features only 2B active parameters, achieving up to 1.7x faster inference than comparable MoEs on-device while matching larger models on key agentic benchmarks.
The new model achieves 13+ tokens/s decoding speed on iPhone 17 Pro Max using Lookahead Expert Scheduling.