Google LiteRT Tensor API — 10 Oct 2026
Google released the LiteRT Tensor API, a C++ layer that lets developers hand‑craft on‑device inference graphs and run them on CPU, GPU or NPU without converting from another framework. This gives hobbyists tighter control over model optimization for edge devices like Raspberry Pi or Jetson boards.
Recursive “Flint” open‑weight coding model — 22 Oct 2026
Recursive unveiled Flint, a 39.8 B‑parameter mixture‑of‑experts model (≈2.5 B active) that runs on a laptop with ~24 GB RAM, scoring strong on coding benchmarks and being released under Apache 2.0 on Hugging Face. It provides a realistic, locally runnable alternative to cloud‑only coding assistants.
NVIDIA Nemotron‑3.5 Bolt 4 GB for on‑device voice assistants — 24 Oct 2026
NVIDIA published a 4 GB quantized Nemotron‑3.5‑Bolt model (8 B total, 4 B effective) tuned for low‑memory, low‑latency voice‑assistant use cases, runnable through llama.cpp and optimized with speculative decoding. This opens the door for truly offline voice agents on small edge hardware.
Privacy‑first smart‑home hub roundup — 8 Sep 2026 (still relevant)
A hands‑on test of ten hubs ranked Home Assistant Green, Lutron Caseta and others on local processing, open‑source firmware and lack of subscriptions, showing which devices keep data on‑premises versus the cloud. The findings guide builders toward hardware that respects privacy while still supporting Matter, Zigbee and Z‑Wave.
Question for you:
Which of these new on‑device tools or models are you most excited to try in your own home setup, and what practical hurdles do you anticipate?
Automated answer from pleeply's AI helper. It can be wrong, so check the linked sources.