Wednesday, Jul 15, 2026
A quiet day in AI infrastructure. The only real movement: a burst of routine llama.cpp server maintenance and one notable Gemma 4 quantization from Intel — nothing that changes what you should build with or pay for.
Also today
Inference & serving
- llama.cpp, 8 releases today (
b10003–b10015) — mostly routine: server gains--cors-*flags (incl. alocalhostshortcut) and a prompt-cache ownership refactor (checkpoints now cleared on prompt clear, state moved intoserver_prompt_cache) are the only user-facing changes; the rest are backend bug fixes — Vulkan/CPU f16 SET_ROWS support, OpenCL dp4a fallback for devices withoutcl_khr_integer_dot_product, DeepseekV4seq_rm/seq_cpfix, a hexagon backend enum-narrowing fix, and a tokenize-tool CLI arg cleanup. No capability jump.
Open weights
- Intel/gemma-4-31B-it-int4-AutoRound-Best — Intel's int4 AutoRound quant of Gemma 4 31B; the one HF upload today from a named lab rather than an individual researcher.