Ripple-65B Preview Achieves New U.S. Open‑Weight Benchmark in Agentic Coding

New State‑of‑the‑Art Performance

Ricursive Intelligence announced Ripple‑65B Preview, a 65‑billion‑parameter language model that now holds the U.S. open‑weight state‑of‑the‑art score on agentic coding benchmarks.

The model attained an 80.52 score on TerminalBench 2.1, surpassing the previous open‑weight leader Reflection Beam 501B, which scored 80.1.

Other large open‑weight competitors fell further behind: Meta Muse Spark Open 816B recorded 76.4, Thinking Machines Inkling 975B achieved 63.8, and NVIDIA Nemotron 3 Ultra 550B managed 56.4.

These results demonstrate that Ripple‑65B can generate high‑quality code with reasoning capabilities comparable to much larger systems.

Efficiency‑Focused Training and Architecture

Despite its strong performance, Ripple‑65B was trained for a total cost of $32 million, roughly one‑tenth the expense of the prior open‑weight SOTA model.

This $32 million figure includes all research and development expenditures, from data mixing and synthetic data generation to the final pre‑training and post‑training runs.

The team avoided purchasing external post‑training data, instead building proprietary synthetic pipelines, quality filters, and coverage metrics to feed the model.

Reinforcement learning was applied using a mixed recipe that incorporated truncated importance sampling, score centering, and a linear length penalty, allowing the model to improve with only 1.4 million rollouts.

By contrast, Reflection Beam required about 70 times more rollouts to reach its benchmark score.

Architecturally, Ripple‑65B employs a dense Transformer design rather than a Mixture‑of‑Experts approach, simplifying both training and serving on NVIDIA hardware.

The model features a “thin‑and‑deep” aspect ratio with a narrow hidden dimension (d_model = 5120) spread across 80 layers, a configuration that supports strong reasoning while keeping the parameter count modest.

During training the model maintained a mean‑field utilization (MFU) of 40 percent, achieved without custom kernels, quantization, or complex sharding strategies.

Its dense architecture also enables low‑latency token generation, making it suitable for real‑time coding assistance.

Deployment, Context Length, and Future Access

Ripple‑65B Preview is pretrained on 12 trillion tokens and can handle contexts up to one million tokens, expanding its ability to process extensive codebases or documentation.

The model is already available through multiple platforms, including Hugging Face, Together AI, GMI Cloud, and OpenRouter, with an additional rollout on Fireworks planned for later today.

Its relatively small parameter envelope allows it to run efficiently on local devices such as NVIDIA DGX Spark, which offer high compute FLOPs with lower memory bandwidth than typical datacenter accelerators.

Ricursive Intelligence positions Ripple‑65B as the general‑purpose variant of the proprietary model that powers its AI chip‑design platform, indicating a broader strategy to leverage the same efficiency gains across domains.

By delivering top‑tier agentic coding performance at a fraction of the traditional cost, the model could lower barriers for organizations seeking advanced code generation capabilities.

Future work will likely explore further scaling of context length and refinement of synthetic RL environments to push the benchmark scores even higher.

Why This Matters: Ripple‑65B shows that high‑performance agentic coding can be achieved with far lower training budgets, potentially democratizing access to cutting‑edge code‑generation AI.