Jan 0.8.5-0.8.7: Local API host check, bundled llama.cpp, no Intel Mac models

Jan, the desktop app for running models locally, shipped three releases in October 2026. **v0.8.5 (Oct 8)** fixes GHSA-x6p8-7cp8-c3p6. Before, a Local API Server bound to `0.0.0.0` accepted any `Host` header. It now returns `403` unless the host is `localhost`, `host.docker.internal`, a loopback or private LAN IP, or an entry in Trusted Hosts. The ID is cited in the release notes; the advisory page itself returned 404 on Oct 20. The same release bundles the llama.cpp engine (llama.cpp 0.5.0, build b11146) in the installer, so Jan no longer downloads a backend. On Windows and Linux, NVIDIA GPUs use a CUDA 13 build. Other GPUs, and NVIDIA GPUs or drivers that cannot run CUDA 13 (such as Pascal GTX 10-series cards), use Vulkan. The CUDA 11/12 download option from 0.8.4 is gone. The engine is built for Apple Silicon only, so llama.cpp models no longer load on Intel Macs; remote and custom providers still work there. The executable is renamed `Jan-Desktop`, and the `jan` CLI is now a separate install. "Remember Conversations Between Sessions" is on by default: each conversation's prompt cache is kept on disk in `llamacpp/thread-cache`, capped by the Conversation Cache Disk Limit setting (8192 MiB). **v0.8.6 (Oct 9)** is a hotfix. It fixes a Linux AppImage crash on model load on some distributions. It ships native Turing code again, so GTX 16xx / RTX 20xx cards on drivers older than R595 can run the CUDA engine. It fixes a Qwen3.8 MTP slowdown on 12 GB-class cards by using one slot when Parallel Sequences is auto. **v0.8.7 (Oct 20)** adds a Shell on/off switch in the chat tool menu and a Max Image Size setting (default 10 MB, range 1 to 100). Its known issues still say local models do not run on Intel Macs.