The Oct 15, 2026 update fixes a bug in the CLAMP operation when inputs are non‑contiguous, improving accuracy on many platforms.
NoteBy OossaPublished by Oossa: 1 min read
The open‑source llama.cpp library released version b11639 on Oct 15, 2026. It adds a new kernel that correctly applies the CLAMP operation to non‑contiguous (strided) tensors. The previous kernel read the wrong elements when the source array wasn’t stored contiguously. The fix introduces a strided version called kernel_clamp_nc, used automatically whenever needed.
Developers can now rely on accurate CLAMP results when running models on hardware that uses strided memory layouts.
Models
The new Kandinsky S2V 1.0 Lite (2B) and Pro (21B) models turn speech and a single photo into 25 FPS talking‑head video, and the code is free under MIT.
Models
The new model scores 86.9 on Artificial Analysis' Speech-to-Speech Index, beating OpenAI and Google, and is now live in the Voice Agent API.
Models
YouTube Music will mark tracks that use generative AI with “Fully Gen AI,” “Partly Gen AI,” or “No Gen AI” labels, starting Oct 14 2026.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
