Local models, the interfaces in front of them, and agents you host yourself.
Open WebUI v0.12.4, published 25 October, rebuilds the Docker image on Python 3.12 and Debian 13, so pip packages inside the container need 3.12 support. It also makes Direct Integrations off actually disable saved tool servers, moves sharing changes into Access settings, and adds lazy skill loading plus visible context compaction.
LiteLLM's 24 October releases sign every Docker image and Helm chart with a new cosign key, remove the eMudhra root CA from its trust bundles, and mark both changes as present in 1.105.4, 1.104.6 and 1.103.8. If you pin the old public key in CI or an admission policy, verification fails until you replace it.
All seven 1.107.0 wheels installed a credential stealer and a remote-code loader that run on every Python start in that environment, not just when LiteLLM runs. PyPI has removed the release and the affected Docker tags are yanked; anyone who installed it needs to rotate credentials, not just uninstall.
The self-hosted AI interface gives direct messages and private channels temporary end-to-end encrypted sessions, encrypted in the browser with MLS and relayed without being saved; the same release removes Google Drive's verification wall for self-hosted installs and lightens Redis streaming.
Salvatore Sanfilippo's engine is deliberately narrow: DeepSeek, GLM and Qwen flash models on Macs, DGX Sparks and Strix Halo boxes, with SSD streaming for hardware that is almost big enough.
Audio scenes that keep track of who said what, ordered failover across local and remote models, structured decision endpoints and signed OCI galleries, in a 15-day window with 243 merged pull requests.
A CLI plus iOS and Android apps drive pi agent sessions in remote sandboxes, with the control plane, sandbox service and both apps in one AGPL repository, installed as a single Compose project.
The self-hosted AI chat's new release lets you interrupt, steer, queue and approve what agents do mid-run, and bundles background tasks, scheduled chats and a refreshed model lineup.
TypeSafe's decision model charges $0.042 per million input tokens with free output; a parody post and two open rewrites show the trick runs on any GGUF model, and a coin-flip test shows where the numbers get shaky.
11.3k stars, one-line install on macOS, Linux or WSL, apps running beside the agent and memory kept as plain files. The modified Apache 2.0 licence is the part to read before you standardise on it.
Anthropic and OpenAI shipped cheaper flagships on Tuesday, and Anthropic's own announcement says Opus 5.5 is the first release since the pacing essay.
fusion-runtime runs speech recognition, an LLM and speech synthesis in one process, and its benchmark section publishes the number that makes it look worse and explains why the stages do not add up.
