Ripple-65B Preview Sets U.S. Open-Weight SOTA on Agentic Coding
Story 16, from Hacker News
Ripple-65B Preview, a new open-weight model, achieves state-of-the-art performance on agentic coding tasks, outperforming larger models with significantly lower training costs.
Key points
- Ripple-65B Preview is a general-purpose variant of a proprietary model used in AI chip design.
- It was pretrained on 12 trillion tokens with up to 1M context length.
- It outperforms models like Reflection Beam 501B, Meta Muse Spark Open 816B, and NVIDIA Nemotron 3 Ultra 550B on TerminalBench 2.1.
- Training cost was $32M, one-tenth of previous open-weight SOTA models.
- The model uses a dense Transformer architecture optimized for NVIDIA hardware.
- A custom model laddering approach improved compute efficiency for code tasks by up to 16x.
Why it matters
This model demonstrates that high performance in agentic coding can be achieved with lower costs and efficient hardware use, offering a practical alternative to larger, more expensive models.
Counterpoint
The model's performance is benchmarked against other open-weight models, but the exact comparison metrics and environments are not fully detailed in the source.
Sources
This summary was written by a language model and may contain errors. Read the original for the full story.