What changed
Ollama 0.40.16 stops silently truncating /v1/responses requests unless truncation is requested, routes `ollama launch claude` to the selected model, lowers Qwen3.8 Flash Next memory use on Apple silicon and adds Ripple Preview plus Windows-on-Arm CUDA 13.5 support.
Why it matters
Silent context loss can change an agent's decision without an explicit failure, while launch routing determines which model receives code and credentials. The release makes both behaviors testable and may materially change local memory headroom for one model family.
What to do
Pin 0.40.16, send an over-limit synthetic Responses request with and without explicit truncation, launch Claude against a selected local model and measure Qwen memory on the target Mac. Confirm errors, model identity, prompt retention and rollback.
Caveat
The signed release is maintainer evidence. Memory savings are model- and hardware-specific, Ripple is a preview, and changed truncation can turn previously completing requests into failures that applications must handle deliberately.