local-ai · review · first seen Nov 10, 10:32 UTC

Ollama releases v0.40.16 with memory and API changes

Ollama shipped v0.40.16 with changes to Apple silicon memory use, Claude Code routing, API responses, and Windows on ARM support. Builders should care most if their workflow depends on predictable model routing or API response handling.

L0 · primary1 source6 claims checked4 min read

What shipped

The release includes Ripple Preview from Ricursive Intelligence, described as a 65B dense model built for agentic coding. The release notes give `ollama run ripple-preview` as the command to run it. That is a shipped model entry, not evidence about how well the model performs on a particular coding task. The notes also say Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon in this version. That figure describes a reduction, not the total memory a given machine will need. [1][2]

The `ollama launch claude` command now always routes Claude Code to the selected model. For a builder using that launch path, the selected model is the stated destination. Separately, the release notes say `ollama launch` explains when a cloud model has been retired. These are changes to launch behavior and messaging; the source does not establish a new Claude Code capability or the outcome of a particular coding workflow. [1]

Two API edges changed. `/v1/responses` respects truncation and no longer silently truncates requests by default. `/v1/models` returns an empty list, rather than null, when no models are installed. Both changes warrant a look at callers that make assumptions about request handling or the no-model response. The release also adds CUDA 13.5 support for Windows on ARM. Support is the claim here; the notes do not provide a performance result for that platform. [1][2][3]

Why it matters if you build

If your Apple silicon setup runs Qwen3.8 Flash Next, the reported memory reduction is a reason to retest the same workload under v0.40.16. About 27 GiB less memory may matter to a constrained setup, but the release gives no absolute memory requirement, hardware configuration, or measurement for your workload. Treat the figure as a release-note claim to check locally, not a capacity guarantee. Ripple Preview is another model to evaluate for agentic coding; its 65B size and stated purpose are not a quality comparison. [1][2]

For a builder connecting Claude Code through `ollama launch claude`, the routing change addresses where the selected model is sent. For an API client, the immediate question is compatibility: check how your caller handles `/v1/responses` requests now that silent default truncation is gone, and make sure its no-model branch accepts an empty list from `/v1/models`. The notes establish the changed behavior, not how any particular client reacts. Review dependent code rather than assuming either change is invisible. [1][2][3]

Windows on ARM builders with a CUDA 13.5 use case now have an explicit support claim to test against their own setup. The release does not specify a throughput gain or say that every Windows on ARM configuration will work. A targeted compatibility check is therefore more useful than a broad migration decision based on the support line alone. [1]

How to try it today

1. Use v0.40.16 and run `ollama run ripple-preview` to evaluate the included coding model.
2. On Apple silicon, rerun your Qwen3.8 Flash Next workload and compare observed memory use with your own previous run.
3. Select a model, then use `ollama launch claude` and check routing in your workflow.
4. Send a representative request to `/v1/responses` and inspect the result against your truncation assumptions.
5. With no models installed, call `/v1/models` and verify that your client handles an empty list.
6. If relevant, test your Windows on ARM CUDA 13.5 setup rather than inferring performance from support alone.

What we checked

Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.

1. shipped L0 · primary

   Ollama v0.40.16 includes Ripple Preview, a 65B dense model for agentic coding.

   per GitHub releases · ollama/ollama (L0)
2. shipped L0 · primary

   Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon in Ollama v0.40.16.

   per GitHub releases · ollama/ollama (L0)
3. shipped L0 · primary

   Ollama v0.40.16 makes `ollama launch claude` route Claude Code to the selected model.

   per GitHub releases · ollama/ollama (L0)
4. shipped L0 · primary

   Ollama v0.40.16 no longer silently truncates `/v1/responses` requests by default.

   per GitHub releases · ollama/ollama (L0)
5. shipped L0 · primary

   Ollama v0.40.16 returns an empty list from `/v1/models` when no models are installed.

   per GitHub releases · ollama/ollama (L0)
6. shipped L0 · primary

   Ollama v0.40.16 adds CUDA 13.5 support for Windows on ARM.

   per GitHub releases · ollama/ollama (L0)

What is not confirmed

The release excerpt does not establish Ripple Preview's coding quality, Qwen3.8 Flash Next's absolute memory requirement or test conditions, the outcome of a particular Claude Code workflow, how a particular API client reacts, or Windows on ARM performance. It confirms the listed changes, not those downstream results. [1][2][3][4][5][6]

Who should act, who should wait

Move now to test if you maintain a client for `/v1/responses` or `/v1/models`, rely on `ollama launch claude` for model selection, or need to assess the reported Apple silicon memory change in your workload. These shipped changes directly touch those paths. If your decision rests on Ripple Preview outperforming your current coding model, a specific memory ceiling, or Windows on ARM speed, wait for your own measurements: the release excerpt establishes none of those outcomes. [1][2][3][4][5][6]

Sources

• v0.40.16 — GitHub releases · ollama/ollama L0 · primary