Ricursive releases Ripple-65B Preview, an open-weight coding model it says cost $32 million to build
The 65-billion-parameter dense model scores 80.5 on Terminal-Bench 2.1, which the chip-design AI lab calls a new best for US open weights.
Why it matters
Ricursive Intelligence trained Ripple from scratch on 12 trillion tokens to power its AI chip-design platform, and released a general-purpose version under Apache 2.0 with up to 1 million tokens of context. It says the model beats much larger US open models, including Reflection Beam 501B, and runs locally on an NVIDIA DGX Spark. The $32 million figure covers its full research cost. Watch independent benchmarks, since the scores are self-reported.
Key facts
- 65 billion parameters, dense, pretrained on 12 trillion tokens
- Up to 1 million tokens of context; Apache 2.0 licence
- 80.52 on Terminal-Bench 2.1 versus 80.1 for Reflection Beam 501B
- Full R&D cost of $32 million, by the company's account
- Available on Hugging Face, Together AI, GMI Cloud and OpenRouter