The open‑source inference library now includes a new fused op for QSA masking, improving performance for certain models.
NoteBy OossaPublished by Oossa: 1 min read
The ggml‑org llama.cpp project released version b11800 on 2026-10-21. The update adds a fused top‑k mask operation (ggml_top_k_mask) that can be used by QSA, DSA and DeepSeek V4 CSA models. The op replaces hand‑crafted masking code with a single node that sets non‑selected cells to –inf. It works per‑head and can handle optional scores, block sizes and extra cells.
Developers can run supported models faster because the library now handles masking in one optimized step.
Models
Canva's new AI creates up to 2K images with detailed layouts and editable text, available to all users via Dream Lab and the API.
Models
The update adds 14 voices, including Hindi and Finnish, and reaches existing generative-voice users without code changes or a price increase.
Models
Uber says its global workforce will use ChatGPT Work after an earlier 1,000-person pilot. The company has not disclosed the rollout’s cost.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
