Ripple-65B Preview delivers strong agentic coding results as a fully Apache 2.0 open US model

Ricursive Intelligence, an AI chip design firm, has released Ripple-65B Preview, an open-weight language model trained on 12 trillion tokens. The company reports an 80.52 score on TerminalBench 2.1, an agentic coding benchmark, which it says is the top result among open US models. The model is published on Hugging Face under the Apache 2.0 license. Ricursive puts total development and training costs, covering all research and development, at $32 million.
Ripple-65B uses a dense Transformer architecture with 65 billion parameters rather than the mixture-of-experts design common among current open models. It is optimized for Nvidia hardware. For evaluation, the company used both Terminus 2 and OpenCode as harnesses across different runs.
The headline result does not carry across every benchmark. Ricursive's figures show Ripple trailing Inkling 2 Small, another recently released western open model, on the other evaluations. The source suggests Ripple is likely easier to deploy on users' own hardware, which fits the dense design: all 65 billion parameters are active for every token, so inference is simpler to run and size but typically heavier per token than a sparse model of comparable total size. A dense model of this scale can plausibly fit on a single multi-GPU node, which this suggests may be part of the pitch.
The Apache 2.0 license is a permissive one, allowing commercial use and modification with attribution. Combined with the published weights, that lowers friction for teams wanting to self-host or fine-tune. How well the reported benchmark translates to real coding workflows will depend on harness choices, which Ricursive discloses only at a high level.
Why it matters: Teams weighing open models for agentic coding now have a US-licensed, self-hostable option with a strong TerminalBench claim and a disclosed cost base. The trade-off is that its advantage appears narrow to coding benchmarks, with a rival ahead elsewhere, so developers should validate on their own tasks rather than assume broad superiority.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Ricursive Intelligence is releasing Ripple-65B Preview, a 65-billion-parameter open-weight language model that the company says achieves the top score among open-source U.S. models on the TerminalBench 2.1 coding benchmark at a total R&D cost of $32 million. On other benchmarks, however, Ripple falls behind the likewise newly released western open model Inkling 2 Small. The article Ripple-65B Preview delivers strong agentic coding results as a fully Apache 2.0 open US model appeared first on The Decoder .