Ollama 0.40.16 Adds Ripple 65B, MLX Path Supports DFlash 2 Speculative Decoding

Ollama released 0.40.16 on November 9, highlighting the addition of Ricursive Intelligence’s Ripple Preview. Developers can launch the model with ollama run ripple-preview; the model page also lists launch commands for Claude Code, OpenCode, and Hermes Agent, allowing existing coding agents to connect to this local model. Release notes

One noteworthy aspect of this integration is its inference path. Ollama says Ripple uses DFlash 2 speculative decoding on the MLX engine to speed up generation. The model page lists model files of approximately 39 to 46 GB. However, file size is not a direct measure of the memory required at runtime; deployments still need to reserve space for the context cache and runtime allocations. The official page does not yet provide speed measurements, hardware conditions, or speculative acceptance rates, so engineering teams will need to measure performance on their own workloads. Ollama model page

Ripple was released on November 4. It is a 65B dense Transformer pretrained on 12 trillion tokens and released under the Apache 2.0 license. Because every token activates all parameters, its total parameter count alone cannot be used to directly estimate its compute cost relative to a mixture-of-experts model. The model card describes an 80-layer architecture with a narrower hidden dimension and a higher proportion of sliding-window attention to control attention costs. Official model card

The long-context figures also require careful interpretation: Ollama’s tag shows a 2M context window, while its descriptive text and the original model card claim support for about 1M tokens. These labels do not establish task quality or local usability at the full context length. Before deployment, teams should verify the actual configuration, memory usage, and ability to retain information over long contexts.

Ricursive reports a score of about 80.5 on Terminal-Bench 2.1, but the model card also acknowledges that results vary with the agent framework, summarization strategy, and sequence length. These BF16 evaluation results do not directly represent the performance of Ollama’s quantized version. For Chinese-speaking technical teams, the next step should be to test Chinese requirement comprehension, tool-call success rates, and end-to-end task time to determine whether the generation gains from speculative decoding translate into practical development efficiency. Evaluation details