Ricursive unveils Ripple-65B, Apache 2.0, and $32M

Ricursive Intelligence launches Ripple-65B Preview, a dense open language model with 65 billion parameters, optimized for Nvidia hardware and released under the Apache 2.0 license on Hugging Face. The company claims a score of 80.52 on TerminalBench 2.1, touted as the best among American open models, following pre-training on 12 trillion tokens. However, other benchmarks place it behind Inkling 2 Small, and it is said to be easier to deploy on user hardware.

$32M Announced and Released Under Apache 2.0

Ricursive Intelligence announces a total cost of $32 million for the development and training of Ripple-65B Preview, covering all research and development expenses. The model is available on Hugging Face under the Apache 2.0 license. It is optimized for Nvidia hardware. According to the company, Ripple-65B Preview could be easier to deploy on user hardware.

Claimed Score on TerminalBench and Peer Comparison

Ricursive Intelligence states that Ripple-65B Preview achieved a score of 80.52 on the coding benchmark TerminalBench 2.1, presented as the best result among American open models. However, other benchmarks show that Ripple does not match Inkling 2 Small, which was recently released.

Dense Architecture and Pre-training on 12 Trillion Tokens

Ripple-65B Preview is based on a dense Transformer architecture with 65 billion parameters, unlike models using a mixture of experts. It was pre-trained on 12 trillion tokens. The model partially utilizes Terminus 2 and partially OpenCode as its framework. Ricursive Intelligence, a company specializing in AI chip design, is behind this open language model.