Neuro Digest for the Week (#111)

Neuro Digest for the Week (#111) LLM - 400B Qwen 3.5 on iPhone 17 Pro — Speed of half a token per second, but there is room for optimization. - GigaChat — Sber released MoE versions (702B and 10B) and shared engineering details of the transition from Dense to MoE. The ultra version surpasses DeepSeek V3 and Qwen3 in mathematics and reasoning. Generative Models - Seedance 2.0 — ByteDance finally released its SOTA video model outside of China. - Google Lyria 3 Pro — The model now generates audio up to 3 minutes, allowing for the creation of full tracks end-to-end. Other - OpenAI shuts down Sora — Neuro-TikTok did not take off, OpenAI decided to focus on more profitable directions. - Dating show with LLM — The team put 12 top models into a dating show format, where they seriously look for love and discuss the size of the context window. Personal - A tale about a German PhD — To insert a publication from an internship into a dissertation, all 13 co-authors need to sign a document stating that the first author did 80% of the work. Read Digest #110