The open‑source LLM runtime now retires old DMA descriptors on Qualcomm Hexagon NPU, preventing NaN outputs.
NoteBy OossaPublished by Oossa: 1 min read
The ggml‑org/llama.cpp project released version b11516 on 2026‑10‑09. The update fixes a bug in the Hexagon NPU path where the DMA ring could overflow, causing stale data and NaN/inf results. The code now drops the oldest descriptor when the ring is full and flushes the queue later. Test cases with large patch‑embed operations now pass on Hexagon hardware.
Developers using llama.cpp on Qualcomm Hexagon devices will get correct model outputs instead of erroneous NaNs.
Models
The company’s new three‑agent Claude‑based reviewer found 73.43% of core‑claim errors, far above the previous best of 14.81%.
Models
Mistral's new trillion‑parameter model scores 38 on an independent index, but costs more per task than competing open models.
Models
OpenAI reported two influence campaigns that used AI‑generated content to plant false stories in real media and blocked the involved accounts.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
