Ollama adds Ripple support and lower Apple-silicon memory use; next release remains a preview
Ollama’s stable 0.40.16 release adds Ripple Preview, Ricursive Intelligence’s 65-billion-parameter coding model, alongside fixes aimed at local inference and agent integrations. The GitHub release is dated November 9. Stable release
The release notes say Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon. That is a reported saving for a specific model and platform, not a claim that every model becomes smaller. The release also changes the Responses-compatible API to respect the requested truncation setting and stop silently truncating by default. Other changes cover routing Claude Code to the selected model, retired-cloud-model messages, Windows-on-ARM support for CUDA 13.5 and a newer llama.cpp backend. Stable release
A separate 0.40.17-rc0 build was published on November 10 and is explicitly labeled a prerelease. Its notes include concurrent generation for the MLX backend, memory-mapped loading for Vulkan integrated GPUs, CLI restart improvements and further backend updates. Those changes should be described as preview-build features until a stable release is confirmed. Release candidate
For local-model users, the immediate significance is compatibility and resource use rather than a new frontier benchmark. This report concerns its packaged availability in Ollama and the surrounding runtime changes, rather than claiming a new Ripple model launch. Neither release was installed or benchmarked for this report, so speed, reliability and total memory requirements remain workload-specific questions.
Sources and citations
Revision history
Published .