Ripple-65B Preview is a general-purpose, open-weight language model developed by Ricursive Intelligence from scratch to support its AI chip-design platform. Pretrained on 12 trillion tokens, it supports context lengths up to one million tokens and is designed for efficient inference, including on local hardware. Its reported TerminalBench 2.1 score of 80.52 makes it the leading U.S. open-weight model on that agentic-coding benchmark, ahead of several models with far more parameters. Ricursive says it built the model for $32 million in total research and development, about one-tenth the estimated cost of the previous open-weight benchmark leader.
The development approach emphasizes compute efficiency rather than scale alone. Ricursive used synthetic post-training data and a carefully tuned supervised fine-tuning and reinforcement-learning process, reaching its benchmark score with 1.4 million RL rollouts - around 70 times fewer than Reflection Beam reportedly used. The model is a dense, 65-billion-parameter Transformer, not a mixture-of-experts system, with an architecture co-designed for NVIDIA hardware and low-latency generation. A sequence of small-scale pretraining experiments refined its architecture and training recipe; at the 4-billion-parameter scale, the resulting changes improved coding compute efficiency by as much as 16 times against an Olmo 3-based baseline. Ripple-65B Preview is available through several model platforms, while an internal version powers Ricursive’s chip-design system.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Investors erased about $2.5 trillion from Nasdaq’s market value as Democratic election victories raised concerns about AI expansion. AI data-center stocks fell amid uncertainty over plans to build more facilities. (reuters.com)
Thinking Machines introduced Inkling 2 Small, a 594-billion-parameter mixture-of-experts model with image input and stronger coding, vision, reasoning, and autonomous-work capabilities. It is available through Tinker with closed weights, with Apache 2.0 weights planned after safety testing. (thinkingmachines.ai)
China plans comprehensive AI rules covering model development, data, computing, and deployment, with safety standards, ethical guidance, and regular reviews. It will ease compliance through transition periods and regulatory sandboxes while requiring companies to disclose risks and incidents and tracking AI’s effects on jobs. (bloomberg.com)
AI can boost productivity as people progress from autocomplete to coding assistants and agents, but overreliance can send them past a peak into low-quality “slop.” The author explains who is most at risk and how to avoid or recover from the habit. (thomasjiralerspong.substack.com)
A North Carolina woman allegedly used ChatGPT to develop a plan to poison her husband, including questions about lethal doses and avoiding detection. Her husband’s lawsuit accuses OpenAI and Sam Altman of failing to stop the chats, while she faces attempted-murder and food-contamination charges. (arstechnica.com)
Z.ai’s upcoming GLM-5.5 models, in 32B, 122B, and 256B sizes, use an internal training pipeline spanning web pages, research papers, books, code, video, and audio. The pipeline includes 1 trillion web pages, 2 billion scientific PDFs, and 500 million code repositories. (twitter.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.