The open‑source llama.cpp library now includes BF16 data type ops and GLU, improving CPU and CUDA performance.
NoteOossaPublished by Oossa: 1 min read
The ggml‑org team released llama.cpp version b11293. It adds BF16 (a half‑precision number format) unary, GLU, binary and scale operations for both CPU and CUDA. The change also updates CUDA kernels to fix a BF16 build issue. The update is available for many platforms, from macOS and Windows to Linux and Android.
Developers can now run llama.cpp models faster on hardware that supports BF16, such as newer GPUs.
Last updated: ·Markdown
Models
Google's new Gemini 4 Argon model, with a 1 million‑token limit and $2 per M input token pricing, is now available to a select group of cybersecurity teams.
Models
Google rolls out reusable “Skills” in Gemini chat and will stop the older Gems feature on Nov 17, 2026 for personal accounts.
Models
Runway Research announced Praxis‑1, an open‑weight robot action model trained on video, with early tests by Noble Machines, Standard Bots and Ultra.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
