TL;DR: Ollama's latest patch updates its bundled llama.cpp runtime for new model support and surfaces installed local models inside Codex's /model picker.
Summary: Ollama shipped v0.40.4, bumping its bundled llama.cpp runtime to b11521 to add support for decision models and EmbeddingGemma 2. The release also integrates locally installed Ollama models into Codex's /model picker and makes inference routes reuse verified manifests and already-loaded models to cut redundant reloads. MLX-side fixes include compacting restored recurrent state after speculative rollback.
Why it matters: Local-first builders get a new embedding model option and tighter Codex CLI integration without extra glue code. Upgrade to try EmbeddingGemma 2 for on-device RAG pipelines and verify the faster inference-route reuse in your own workloads.
Source: github_releases