Ricursive Intelligence, a full-stack AI lab, has released Ripple-65B Preview, an open-weight language model trained from scratch to achieve state-of-the-art performance on agentic coding tasks. The model scores 80.52 on TerminalBench 2.1, according to the company, surpassing significantly larger competitors including Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4).
Ripple-65B Preview was pretrained on 12 trillion tokens and supports up to 1M context length. The model is available on Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks.
A key distinguishing factor is cost efficiency. Ricursive developed and trained the model for $32 million in full R&D spending—including pretraining, data mixing, evaluation, synthetic data generation, and posttraining—which the company estimates is one-tenth the cost of the previous open-weight agentic coding SOTA. This contrasts sharply with broader industry trends: OpenAI’s Sam Altman stated in 2023 that GPT-4 training exceeded $100 million, while Anthropic’s Dario Amodei projected frontier models could cost up to $1 billion to train, with pretraining costs potentially reaching $10-100 billion by 2026-2027.
The model uses a dense Transformer architecture rather than Mixture-of-Experts, designed for efficient training and serving on NVIDIA hardware. Ricursive optimized the architecture with a narrow model dimension (5120) and 80 layers, along with features including gated attention, global NoPE, and value residuals. The company maintained 40% model FLOPs utilization during training using simple sharding without custom kernels or quantization.
Ripple’s reinforcement learning approach differed from typical industry methods. Rather than brute-force RL scaling, the company iterated carefully on supervised fine-tuning and RL recipes using internally developed synthetic data pipelines. The model reached its TerminalBench 2.1 score with only 1.4 million rollouts—roughly 70x fewer than Reflection Beam, according to Ricursive.
The company also refined its pretraining recipe through extensive model laddering experiments, testing architectural interventions at small scale with a focus on coding capabilities. These iterations, including switches to pre+post normalization, sigmoid gated attention, and global NoPE, compounded to improve compute efficiency on code by up to 16x at 4B model scale compared to a baseline, according to Ricursive’s measurements.
Key facts
- Ripple-65B Preview achieves 80.52 on TerminalBench 2.1, surpassing larger open-weight models
- Developed for $32 million in total R&D cost, estimated at one-tenth the prior open-weight agentic coding SOTA cost
- Model pretrained on 12 trillion tokens with up to 1M context length support
- Uses dense Transformer architecture rather than Mixture-of-Experts for efficiency
- Reached state-of-the-art score with 1.4M RL rollouts, roughly 70x fewer than Reflection Beam
- Available on Hugging Face, Together AI, GMI Cloud, OpenRouter, and Fireworks