Ricursive Intelligence has released a preview of Ripple-65B, an open language model trained on 12 trillion tokens. The system achieved a score of 80.52 on the TerminalBench 2.1 coding benchmark, outperforming other open US models in that specific test. Ricursive states that the total development and training costs reached just $32 million. The model relies partly on Terminus 2 and partly on OpenCode as its harness. While it does not match the performance of the newly released Inkling 2 Small on several metrics, it offers a distinct advantage for local deployment.
The model uses a dense Transformer architecture with 65 billion parameters instead of the more common mixture-of-experts approach. It is specifically optimized for Nvidia hardware. This design choice means users can run the model on their own machines without relying on external cloud services. The code is available on Hugging Face under the Apache 2.0 license.
* Dense architecture avoids the overhead of mixture-of-experts routing.
* Optimized specifically for Nvidia hardware infrastructure.
* Fully open under Apache 2.0 licensing terms.
![Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL [R]](https://ai-maestro.online/wp-content/uploads/2026/05/masked-diffusion-language-models-are-strong-and-steerable-te-768x768.jpg)