The Sep 30, 2026 release tweaks Vulkan tile selection, cutting wasted work in Mixture‑of‑Experts models.
NoteOossa1 min read
The open‑source llama.cpp library added a Vulkan fix for MoE‑aware matrix multiplication. The change adjusts how the code picks tile sizes for the mat_mul_id operation, matching the true number of active rows per expert. In a test on a 30‑billion‑parameter model, the fix reduced idle GPU work that had taken 55% of the run time.
Less idle GPU time means faster inference for large MoE models running on Vulkan‑compatible hardware.
Models
OpenAI’s new GPT‑6.1 Sol keeps the $2‑$10 per‑million‑token rate while claiming stronger performance, undercutting Anthropic’s Claude and DeepSeek models.
Models
The new v4 and v4 Turbo voices cost $0.08 and $0.04 per 1,000 characters, with a promotional cut to $0.022 and $0.011 per 1,000 until Oct 12.
Models
Oracle’s new Fusion Claw runtime lets its Fusion apps handle multi‑step jobs like staffing and accounting, using AI for reasoning while keeping calculations outside the model.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
| # | Source | Outlet | Date | Key takeaway |
|---|---|---|---|---|
| 1 | ggml-org/llama.cpp b11265 ↗ | llama.cpp | Sep 29, 2026 | <details open> vulkan: MOE aware mat_mul_id tile selection (#29182) mut_mul_id selected its matmul tile with total token count. |
1 sources
