Ollama 0.40

Apply these rules to Ollama serving, model imports, Modelfiles, CLI use, and native or compatibility API integrations. Review applications and SDKs connected to Ollama under their own trust boundaries.

Release and deployment integrity

• Follow Ollama's latest stable release for security fixes; look up the exact release in the official release index instead of trusting memory. Never loosen an existing deployment constraint just to complete an installation.
• Verify official download checksums or container digests, keep image references current through automation, and audit the image, bundled native engines, and host separately in CI.
• Build custom binaries and native runners from the same reviewed source tag with its declared Go and native toolchain requirements; test the matched runner payload and GPU driver together. Prebuilt installations do not need a separate Go runtime.
• Keep the host operating system and accelerator drivers supported by the selected release. Verify the effective inference backend after upgrades: on Apple Silicon, supported architectures use MLX by default, which changes the native parser and device surface exercised by a model.
• Test the approved model fleet on each enabled backend after changing runners, drivers, or model formats; reject a fallback to an unreviewed backend or a silent change in model capabilities.

Listener and API access

• Keep OLLAMA_HOST on the default loopback listener unless remote access is needed. The official container sets OLLAMA_HOST=0.0.0.0:11434; bind any published container port to the intended interface