The open‑source LLaMA inference library now handles BF16 math, improving performance on several backends.
NoteBy OossaPublished by Oossa: 1 min read
The ggml‑org team released llama.cpp version b12202. It adds support for the BF16 (brain‑floating‑point 16‑bit) precision in key matrix multiplications. The change mirrors the reference implementation and is now accepted by CUDA, Vulkan, OpenCL, Hexagon and other backends. Tests for flash‑attention now run with BF16 alongside the existing F32 mode.
Developers can run LLaMA models faster on hardware that prefers BF16, such as newer GPUs.
Models
The model is available through the company’s Tinker service. Thinking Machines says it can handle coding, research and business tasks, and plans to release its weights.
Models
The 65-billion-parameter model is available under Apache 2.0. Ricursive reports a strong coding benchmark score, but results vary across tests.
Models
A North Carolina man alleges ChatGPT helped his wife plan to poison him. OpenAI says it banned her account and shared information with police.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
