Ricursive Intelligence releases Ripple-65B Preview
The new open language model offers a dense architecture for companies requiring local deployment.
On 6 November 2026, Matthias Bastian reported for The-decoder that Ricursive Intelligence released a new open language model named Ripple-65B Preview. This model was pretrained on 12 trillion tokens and carries an Apache 2.0 license. The company claims the model achieved a score of 80.52 on the TerminalBench 2.1 coding benchmark. Ricursive Intelligence stated that total development and training costs reached $32 million. While the model trails the Inkling 2 Small in some benchmarks, it uses a dense Transformer architecture instead of a mixture-of-experts approach.
Why it matters
The decision to use a dense Transformer architecture rather than a mixture-of-experts model affects your infrastructure requirements. While Inkling 2 Small shows higher benchmark scores, Ripple-65B Preview is likely easier to deploy on your own hardware. For Swiss companies prioritising data sovereignty, the Apache 2.0 licence allows you to host the model locally without external dependencies.
- You can run the model on private servers to maintain control over proprietary code.
- The $32 million development cost suggests a more accessible price point for open-source alternatives.
- Hardware optimisation for Nvidia simplifies the integration into existing local setups.
What to watch
You should assess your current Nvidia hardware capacity before committing to a dense 65 billion parameter model. Since Ripple-65B Preview is optimised for Nvidia, your existing compute resources will determine the efficiency of local deployment. Compare the specific performance gains in your own coding workflows against the higher benchmark scores of mixture-of-experts alternatives.
Within a year, dense models will lose their deployment advantage. This happens if mixture-of-experts architectures achieve the same memory efficiency as Ripple-65B's current architecture.
Key takeaways
- Dense architectures may simplify local hardware deployment compared to mixture-of-experts models.
- The Apache 2.0 licence supports data sovereignty through local hosting.
- Deployment efficiency depends entirely on your existing Nvidia hardware capacity.
Frequently asked
How does the model's performance compare to its competitors?
It trails the Inkling 2 Small in certain benchmarks despite its specific architecture.
Will this model remain the most efficient choice for local hosting?
It depends on whether mixture-of-experts architectures improve their memory efficiency within the next year.