Who contributed to Local LLMs on July 24, 2026?

12 developers shipped this update, including Jesse Gross, github-actions[bot], yuyue0225sc, davidjpyu, izhuhaoran, chaeminlim-mb, mgoin, and rainj-me, and 4 more.

What were the notable Local LLMs updates?

x/create: quantize lm_head at 8-bit in the requested family, x/create: quantize lm_head at 8-bit in the requested family, and b10103.

LLAMA.CPP SHIPS FIVE RELEASES IN 24 HOURS, FIXES SPECULATIVE DECODING ACROSS THE STACK

The inference stack is consolidating around speculative decoding, and the bugs surfacing there are getting squashed across three major projects simultaneously.

llama.cpp burned through five releases in the last day, each one tightening the hardware-specific kernels that matter at scale: CUDA's NVFP4 W4A4 quantization [4] now fuses per-channel amax and quantization to cut overhead, Hexagon activation ops [5] switched to striped DMA for non-contiguous tensors, and Metal got f16 support for leaky relu [3]. Meanwhile vLLM fixed a critical bug where speculative decoding silently broke grammar parsing by missing reasoning boundaries [7], a stop-string correctness bug that only surfaces under speculative decoding [8], and an OOM bug in rejection sampling that chunking now solves [9]. SGLang hit platform expansion: Intel XPU support shipped [11] and the Rust build switched to a workspace model [10]. The pattern is clear: speculative decoding is now the default path, and the three projects are racing to make it bulletproof. Ollama fixed an asymmetry in lm_head quantization [1] where untied heads now follow the same 8-bit rules as tied embedding models, bringing quality closer to bf16 baseline [2]. llama.cpp's speculative draft inference also got smarter: it now infers the draft type from repo sidecars when available [6], eliminating a manual flag. Across all desks, 76 commits and 76 PRs landed in the last 24 hours, with five releases shipped.