ollama v0.40.16

Ollama 0.40.16 adds Ripple Preview, so the open-weight coding model that Ricursive Intelligence announced days earlier can now be downloaded and run on a local machine with one command, ollama run ripple-preview (release notes). Ollama is a tool for running language models on your own hardware. Ripple is a 65-billion-parameter dense model, meaning every weight is used for every token rather than a subset as in mixture-of-experts designs, and Ricursive built it for agentic coding, where a model works through multi-step terminal tasks rather than answering a single prompt. Ricursive describes it as a general-purpose version of the model behind its chip-design platform.

The Ollama listing gives the practical details. Its downloads range from 39 to 46 GB, which gives a sense of the memory a machine needs to load it, though Ollama states no minimum hardware. It is tagged for tools and thinking, and on Ollama's MLX engine for Apple silicon it uses a technique called speculative decoding to generate faster. Two details do not line up with other sources: the listing says the model is licensed under Apache 2.0 with commercial use allowed, while Ricursive's announcement does not state a license, and the listing's tags show a 2-million-token context window while its own description and Ricursive's page say up to 1 million. Anyone planning commercial use should confirm the license on the Hugging Face model card.

The reason to try Ripple is Ricursive's claim, covered in earlier Intel on Ripple-65B Preview, that it scores 80.52 on TerminalBench 2.1, narrowly ahead of Reflection AI's 501-billion-parameter Beam at 80.1 and above other open-weight models several times its size. That figure is Ricursive's own, its page does not describe how the benchmark was run, and the model is labeled a preview. Running it locally through Ollama is a way to test the claim on your own tasks, but this release adds no new evidence about its quality.

The same release fixes a quieter problem for developers who call Ollama through its OpenAI-compatible interface. Its /v1/responses endpoint had been cutting long inputs short by default without saying so; it now leaves input intact unless the request asks for truncation. The release notes also say ollama launch claude now always routes Claude Code to the model the user selected, and that Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon, without explaining how.