Ricursive Intelligence has released Ripple-65B Preview, a general-purpose language model developed from scratch to support its AI chip-design platform. The company claims it achieves a U.S. open-weight state-of-the-art score of 80.52 on TerminalBench 2.1, surpassing much larger models including Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B, and NVIDIA Nemotron 3 Ultra 550B.
Ripple-65B was pretrained on 12 trillion tokens with up to a one-million-token context window. Its dense, 65-billion-parameter architecture is designed for efficient training, low-latency generation, and local deployment on systems such as NVIDIA DGX Spark. The model is available through Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks planned to follow.
Ricursive reports that the full research and development effort cost $32 million, including data development, evaluations, infrastructure debugging, synthetic-data generation, pretraining, and posttraining. The company emphasizes compute-efficient supervised fine-tuning and reinforcement learning rather than large-scale brute-force methods. Ripple reached its benchmark score using 1.4 million rollouts—approximately 70 times fewer than the prior leading model, according to Ricursive.
The model uses a modernized dense Transformer rather than a mixture-of-experts design, achieving roughly 40% model FLOPs utilization without custom kernels or quantization. Architectural changes included gated attention, global NoPE, value residuals, and a thin, deep structure. Ladder experiments reportedly improved coding compute efficiency by up to 16 times at smaller scale. Ricursive says the project advances its broader goal of co-optimizing AI models, software, and hardware.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Sam Altman, Dario Amodei, Ricursive Intelligence, and OpenAI
🤖 AI prefill
📰What
Ricursive Intelligence has released Ripple-65B Preview, a general-purpose language model developed from scratch to support its AI chip-design platform.
🤖 AI prefill
📅When
2023 and 2026 to 2027
🤖 AI prefill
📍Where
United States and Colossus
🤖 AI prefill
💡Why
Centers on Ripple-65B Preview, agentic coding, and U.S. open-weight state-of-the-art
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
Manipulation Score: LowHow much this article uses emotional framing vs. straight reporting
Loaded Language1025%
Absolutist815%
Threat/Panic225%
Us vs Them415%
Engagement Bait1820%
Manipulation score look right?
Factual Claims (10)
mediumtechnology
Ripple-65B Preview is a general-purpose model pretrained on 12 trillion tokens with support for up to 1M context length.
hightechnology
Ripple-65B Preview achieved a score of 80.52 on TerminalBench 2.1, described as U.S. open-weight state-of-the-art.
hightechnology
Ripple-65B Preview surpassed larger models including Reflection Beam 501B, Meta Muse Spark Open 816B, Thinking Machines Inkling 975B, and NVIDIA Nemotron 3 Ultra 550B.
higheconomic
The model was developed and trained for $32 million in full R&D costs, including pretraining, posttraining, data generation, evaluation, and debugging.
higheconomic
The $32 million cost is estimated to be one-tenth the cost of the previous open-weight state-of-the-art model on agentic coding.
hightechnology
Ripple reached a TerminalBench 2.1 score of 80.5 with 1.4 million rollouts, approximately 70 times fewer than Reflection Beam.
mediumtechnology
The model maintained approximately 40% model FLOPs utilization during pretraining and posttraining with simple sharding, no custom kernels, and no quantization.
lowtechnology
The model uses a dense Transformer architecture with a 5,120-dimensional model width and 80 total layers.
hightechnology
Changes to the model ladder improved code-training compute efficiency by up to 16 times at the 4B model scale.
mediumeconomic
OpenAI CEO Sam Altman said GPT-4 training costs exceeded $100 million, while Dario Amodei said frontier models can cost up to $1 billion and projected $10B-$100B costs by 2026-2027.
Comments
Login to view and post comments.