The new Ollama version reduces a specific model's memory footprint on Apple silicon and changes the Responses API's default silent input truncation. Local model developers can focus on memory headroom, while service integrators can focus on preserving input. Support for graphics computation on Windows ARM devices has also been added.
1. Lower memory usage on Apple silicon
In v0.40.16, released on November 9, Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon, according to the company. The announced reduction in memory needed to run the model is a key change to consider when choosing an environment.
Memory is the workspace that holds a model, so lower usage can ease capacity constraints on the same device. Users balancing model execution with other tasks can assess these savings separately from speed improvements.
The announcement does not specify the test device, model settings, or total memory usage before and after the reduction. The 27 GiB figure alone therefore cannot determine whether a specific device can run the model.
2. The Responses API no longer silently truncates input by default
/v1/responses now respects truncation, the setting that controls whether input is truncated, and does not silently truncate input by default. Changing the behavior that silently removed parts of user requests makes it easier to review a service's intended input handling policy.
For developers, the benefit is less unintended loss of input. In return, services need to review input handling previously left to the defaults. The focus shifts from the convenience of automatic truncation to directly managing request handling policies.
3. CUDA 13.5 support added on Windows ARM
This version supports CUDA 13.5 on Windows on ARM. This matters to users who want to run models using graphics computation in ARM-based Windows environments. While the memory savings concern the model and memory capacity, this change concerns platform support.
The release assets include ollama-windows-arm64.zip for Windows ARM64. Users in that environment can review this package alongside the new CUDA support.
4. Integrated services should check long inputs and empty model lists
/v1/models returns an empty list rather than null when no models are installed. Clients parsing the response should ensure they handle the absence of models as an empty list.
Services preparing to update should test requests with long inputs and cases with no installed models. Checking how input is handled and what appears when the model list is empty turns these API changes into practical integration checks.
5. Ripple Preview introduced
The release also introduces Ripple Preview from Ricursive Intelligence. It is a 65-billion-parameter model designed for agents performing development tasks, with a dense architecture that uses all parameters rather than a subset. The following command from the release provides a starting point for users trying the new model.
ollama run ripple-previewOllama v0.40.16 improves memory usage and platform options for local execution while prompting integrated services to revisit their input handling policies.