Ollama shipped a long run of 0.40.x releases through October and early November 2026. Taken together, the changelog shows a project that is moving beyond simply running chat models and towards being a hub for local and cloud-assisted AI workflows.
Performance and model support
Apple Silicon users get continued MLX work. A late-October release candidate preserved small-image detail for the compact vision towers of Gemma 4, and added support for DFlash2, a speculative-decoding approach, alongside another bump of the bundled llama.cpp version. Earlier builds added new GGUF tensor types such as q2_0 and fixed a Windows problem where orphaned runner processes could linger.
Models for coding agents
Release 0.40.9 introduced Flint, a coding model from Recursive aimed at the kind of repository-wide work that agents do: understanding unfamiliar code, tracing bugs, implementing features and adding tests. Another build included a draft model for speculative decoding to speed up generation.
Speech and agent launchers
Ollama is also expanding into speech, with a voice catalog added in 0.40.11. On the agent side, the ollama launch command for tools such as Codex and Claude was polished, with spurious warnings removed, and installed local models now appear in Codex’s model picker. That makes it easy to point a coding agent at a model running on your own machine.
Upgrade notes
Version 0.40.14 became the stable release on November 3, 2026. One migration detail to remember: when models are converted to a newer on-disk format, Ollama keeps backups so that downgrades are safe, which temporarily uses extra disk space. Check free storage before updating a machine with many large models.