An artificial intelligence chip design firm has entered the open-weight coding race with a release that tests the baseline cost of enterprise-grade developer automation. According to Ricursive Intelligence, total development and training costs amounted to just $32 million, an expenditure covering all of the company's research and development expenses for the project.
The model is available on Hugging Face and is distributed under the Apache 2.0 license, trained on a dataset consisting of 12 trillion tokens. On the evaluation front, Ripple-65B Preview achieved a score of 80.52 on TerminalBench 2.1.
The model partly uses Terminus 2 and partly OpenCode as its harness. While separate benchmark results show Ripple-65B Preview does not match the newly released Inkling 2 Small, as Ricursive Intelligence stated, the system is likely easier to deploy on users' own hardware.
Engineering choices behind the project reflect a targeted push toward operational predictability on proprietary clusters. Ricursive Intelligence built Ripple-65B Preview on a dense Transformer architecture with 65 billion parameters instead of the more common mixture-of-experts approach, optimizing the setup specifically for Nvidia hardware. Deploying a 65-billion-parameter dense framework carries distinct trade-offs for corporate balance sheets: maintaining continuous serving without sparse routing concentrates high memory demands on existing infrastructure, leaving engineering teams to measure self-hosted inference expenses against the operational overhead of third-party APIs.
Side by side
Where competitors leaned on sparse mixture-of-experts setups to cut operational friction, Ricursive deployed a dense 65-billion-parameter engine tailored directly to Nvidia hardware. Dropping the expert routing scheme makes proprietary on-premise clusters far easier to manage, but it shifts the cost burden directly onto local video memory. Relying on a dense network forces companies to absorb high local hardware expenses just to bypass third-party APIs.