Ollama 0.33

Apply these rules to Ollama 0.33 serving, model imports, Modelfiles, CLI use, and local or cloud API integrations. Treat client SDKs, gateways, and applications launched through Ollama as separate trust boundaries.

Version, platform, and release integrity

• Install Ollama from the official assets of its newest release, verify the published SHA-256 checksums, and reference container images by a digest that an update bot keeps current; a verified Git commit is not a signature on every release binary. Ollama ships fixes in new releases and asks users to update regularly, so take any exact version from the Ollama releases page rather than memory, leaving a placeholder that fails the build when you can't check. The 0.33 in this prompt's title names the release line these rules cover, and 0.33.2 is the release they were checked against, not a version to install.
• Use maintained, patched operating systems meeting the release requirements: macOS 14 or newer, Windows 10 22H2 or newer, or a compatible Linux deployment; prebuilt Ollama does not require a separately installed Go runtime.
• Match the accelerator payload to supported hardware and drivers: NVIDIA compute capability 5.0+, drivers 550+ on Linux or 551.61+ on Windows, and 570+ for capability 5.0–6.2; use a ROCm v7-compatible stack for AMD ROCm acceleration.
• Build custom binaries with the tag's Go 1.26.0 requirement, CMake 3.24+, and matching native payloads; rebuild and test native components together instead of mixing runner libraries from another release.
• Check current Ollama and bundled native-engine advisories