jul 27, 2026
7 links from the engineering internet.
llama.cpp b10152 keeps mtp front layers on the gpu
the b10152 build counts nextn (mtp) speculative-decoding blocks in n_gpu_layers so the front transformer layers stay on the gpu instead of spilling to cpu, keeping multi-token prediction fast.
ollama 0.32.5-rc0 refreshes the mlx inference engine
the 0.32.5-rc0 pre-release bumps ollama's mlx engine, updating the apple-silicon inference backend ahead of the next stable ollama build.
n8n 2.31.7 stops loading execution data during credentials audit
the 2.31.7 patch stops the credentials security audit from loading full execution data, and signs s3 object paths with strict rfc 3986 encoding so keys with special characters authenticate correctly.
opencode 1.18.7 fixes the macos fullscreen titlebar inset
the 1.18.7 desktop release drops an extra titlebar inset in macos fullscreen, stops removed shadowed commands from reappearing in the command palette, and adds scrolling to a long project-selector dropdown.
prisma 7.9.1 patches a transitive dependency cve in @prisma/dev
the 7.9.1 patch bumps a vulnerable transitive dependency reached through @prisma/dev in the prisma cli. maintainers say it does not actually affect the cli but recommend upgrading to silence security scanners.
astro 7.1.4 fixes duplicate css and client:only scoped styles
the astro 7.1.4 patch stops duplicate css being emitted when prerendered and server-rendered pages share styles, restores scoped styles for child components inside client:only islands, and fixes tsconfig path aliases in astro.config.ts.
code-server 4.130.0 syncs the browser ide to vs code 1.130
code-server 4.130.0 rebases the browser-based vs code distribution onto upstream vs code 1.130.0, pulling the latest editor release into remote and self-hosted ide setups.