Ollama v0.40.16 adds Ripple Preview, a 65B coding model
The release also cuts Qwen3.8 Flash Next memory use on Apple silicon and fixes two OpenAI-compatible API endpoints.
What landed on 9 November
Ollama released v0.40.16 on 9 November, and the headline addition is Ollama Ripple Preview, a 65-billion-parameter dense model from Ricursive Intelligence that the notes describe as built for agentic coding. You pull it the same way as any other model, with ollama run ripple-preview. The tag also carries a memory improvement for Qwen3.8 Flash Next on Apple silicon, three changes around the ollama launch command, two fixes to the OpenAI-compatible HTTP endpoints and CUDA 13.5 support for Windows on ARM. The v0.40.16 release notes on GitHub list 15 commits since v0.40.15.
That is the full extent of what the announcement says about Ripple Preview. There are no benchmark numbers, no stated context window, no license and no list of quantisations or download sizes. The name and the size class are the whole story for now.
The unglamorous fixes that affect running code
For anyone whose app talks to Ollama locally, several of the smaller lines matter more than the new model.
Qwen3.8 Flash Next now uses about 27 GiB less memory on Apple silicon. On a laptop that is often the difference between a model that loads and one that swaps to disk.
ollama launch claude now always routes Claude Code to the model you selected. Previously the routing behaviour was apparently not guaranteed, which is the kind of thing that silently sends your prompts somewhere you did not intend. In the same area, ollama launch now explains when a cloud model has been retired, which turns a confusing connection failure into a readable message.
Two endpoint fixes land on the OpenAI-compatible side of the server. /v1/responses now respects truncation and no longer truncates silently by default, so output limits are surfaced instead of quietly dropping content. /v1/models returns an empty list rather than null when no models are installed, which removes a common crash for client code that iterates over the response without a null check.
The release also bumps the bundled llama.cpp and adds CUDA 13.5 support for Windows on ARM.
What it changes for people building apps
Agentic coding is a different workload from chat. It means long, multi-step loops where the model reads files, proposes edits, runs something and reads the result. A 65B dense model aimed at that job is a signal about where local inference is being pointed, though without published evaluations there is nothing concrete to compare.
The practical value this week sits in the plumbing. If your app uses Ollama through the OpenAI-compatible endpoints, the truncation and /v1/models changes remove two failure modes that are hard to debug from the app side. If you run on Apple silicon, the Qwen memory reduction is measurable and free. If you are on Windows on ARM with an Nvidia GPU, the CUDA 13.5 support is new ground rather than a refinement.
Limits and open questions
The announcement is a changelog, so it answers very little. Whether Ripple Preview runs comfortably on consumer hardware is unstated, and a 65B dense model is a substantial download in any quantisation. There is no word on tool-calling behaviour, context length or how it performs against other coding models. "Flash Next" is given without explanation of how it relates to other Qwen releases. The truncation change to /v1/responses is described but not documented in detail, so it is worth testing your own requests before assuming identical output. The release is dated 9 November without a year in the notes themselves.
For AI app builders
Most people shipping small web apps will not run a 65B model locally, but the model layer keeps moving and the choice of which model sits behind your app is now a configuration decision rather than a rewrite. Draftship takes an app description and builds and hosts it, with a built-in database and sign-in, and paid plans include the option to choose which AI model powers the build chat. Nothing in this Ollama release changes that, and Draftship is not affiliated with Ollama or Ricursive Intelligence. The relevant takeaway is narrower: local inference tooling is maturing on the operational details, and those details are usually what break an app in production.
Primary sources
- v0.40.16 · Ollama releases, Nov 9, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.