Ollama 0.40.16 ships Ripple-65B Preview, a dense Apache-2.0 coding model
Three Ollama releases in a week add Ricursive's 65 B dense model, MiMo V2.6, YaRN long context for Qwen3, and stop OpenAI-compatible endpoints overriding a model's sampler defaults.
Ollama published v0.40.16 on 9 November 2026, six days after v0.40.14 and v0.40.15 landed together on 3 November. The headline model is Ripple Preview from Ricursive Intelligence, which the release describes as a 65 B dense model aimed at agentic coding. The Hugging Face model card gives 65 B total and active parameters (there is no expert routing), a context of up to 1,048,576 tokens, an Apache-2.0 licence and pretraining on 12 T tokens of mostly English text and code, but no release date. The card says post-training was relatively light and aimed at terminal-based coding, and that the model fits on an NVIDIA DGX Spark. The vendor reports 80.5 on Terminal-Bench 2.1 and 76.5 on SWE-bench Verified.
Elsewhere in 0.40.16, Qwen3.8 Flash Next uses about 27 GiB less memory on Apple silicon, /v1/responses now honours the truncation setting instead of truncating silently, /v1/models returns an empty list rather than null when nothing is installed, and Windows on ARM gains CUDA 13.5 support. The 0.40.15 release added Xiaomi’s MiMo V2.6, a 9 B model for code and agent work, the AutoTrust decision models in 9 B, 35 B and 27 B vision sizes, YaRN long-context scaling for Qwen3 and Qwen3.5, and concurrent requests for recurrent and vision GGUF models. The ollama launch command gained Kimi Code CLI and now fills in Pi’s model list automatically.
Why it matters
A dense 65 B model at 1 M context is a different serving proposition from the recent 500 B-plus mixture-of-experts releases: more compute per token, but no routing and a memory footprint that one workstation can hold. The Terminal-Bench figure is self-reported, and the card itself says agentic scores are hard to reproduce across harnesses, so run your own tasks before switching a coding agent to it. The YaRN change means Qwen3 models can be pushed past their trained context, which trades accuracy for length, so test retrieval at the lengths you intend to use rather than trusting the configured maximum.