Ricursive Releases Large-Scale LLM 「Ripple-65B」

US startup Ricursive Intelligence has released the open-weight language model 「Ripple-65B Preview」 with 65 billion parameters. The company is making it available for free under the Apache 2.0 license, opening it to a wide range of applications including commercial use. The total research and development cost for development is reportedly $32 million (approximately 4.8 billion yen).

In the AI model development competition, the open-source community and US-based developers are gradually increasing their presence. Compared to commercial major models, open models that are easier to run in-house environments and have fewer license restrictions are becoming realistic options for enterprises and research institutions. Within this trend, Ripple-65B is positioned as an open model originating from the US.

In terms of performance, Ricursive announced that it achieved the highest score among open-source US models on the coding benchmark 「TerminalBench 2.1」. TerminalBench measures the ability of AI to execute and manipulate code in actual terminal environments, serving as a metric for evaluating coding capabilities as so-called 「AI agents」. Strong performance in this domain indicates the potential for this model to be well-suited for software development automation applications.

On the other hand, results on other benchmarks show that it falls short of another Western open model 「Inkling 2 Small」 released around the same time. This demonstrates that the superiority on TerminalBench may be limited to specific use cases, and careful evaluation aligned with intended purposes is essential when selecting models. The difficulty of judging overall performance based on a single benchmark result is once again highlighted.

The release under the Apache 2.0 license is also noteworthy. This license allows free use, modification, and redistribution for both commercial and non-commercial purposes, and falls into the category of open-source licenses with fewer restrictions. It facilitates developers and enterprises using the model as-is or integrating and improving it within their own systems, providing an advantage in terms of adoption and dissemination.

In the open LLM (large language model) domain, well-known entities like Meta's Llama series remain ahead, but cases of emerging models like Ripple-65B demonstrating competitiveness in specific areas are increasing. As attention to AI agent-driven coding automation grows, the development approach emphasizing practical capabilities including terminal operations aligns with market needs in certain respects. Going forward, improvements in other benchmark domains and actual adoption in real development environments will be key factors determining this model's evaluation.