llama.cpp adds MTP and halves indexer memory for Qwen Flash Next

A LocalLLaMA poster reports using Qwen3.8 Flash Next in "DeepSeek Harness" to repair the Dash2Dock Lite GNOME extension after a major GNOME update broke it, restoring their dock in about an hour.

Arize AI's Fuad Ali walks through a Wonder Toys shopping assistant that returned out-of-budget products because its price filter was broken, using the failure to demo a loop: collect the trace, investigate the code, propose a fix, and test it.

The tool targets orbit-style videos from video models such as Minimax H3 in ComfyUI as input for Gaussian Splatting, skipping the usual COLMAP, trainer, and CUDA setup.

Thor Schaeff, who works on developer experience at Google DeepMind, runs live demos of Gemini's real-time and media models, building an illustrated story with Gemini Live. The talk covers agents that hear, see, and answer in the user's language in real time.

AWS's Elizabeth Fuentes Leone and Sandhya Subramani build a customer service agent that looks up accounts, checks orders, and handles refunds using Strands Agents. The talk focuses on controlling what happens between those steps and keeping that control in production.