Ripple-65B-Preview: Agentic coding performance that beats 550B MoE with 65B

Ripple-65B Preview, released by Ricursive Intelligence, is a 65B parameter dense Transformer model. It was pre-trained on 12 trillion tokens and supports a maximum context length of 1M. It adopts a modernized dense architecture rather than an MoE structure, designed for efficient training and serving on NVIDIA hardware.

It stands out in the field of agentic coding. It scored 80.52 on TerminalBench 2.1, outperforming US open-weight models with more than 10 times the total parameters, such as Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4). It also dominates competing models with a score of 79.1 on SWE-bench Pro v2.0 Hard.

It also has strengths in training efficiency. It improved code Compute Efficiency by 16 times compared to Olmo 3, showing a 55x advantage when scaled up to 55B. It maintained 40% MFU when trained in BF16 precision, and stable inference is possible with simple sharding alone, without custom kernels or quantization.

It provides NVFP4 checkpoints to enable operation on consumer hardware such as the NVIDIA DGX Spark. In a data center GPU environment, it achieves a fast generation speed of 250 tokens per second based on a 2xB200 batch of 8. It is distributed under the Apache 2.0 license, making it easy to use for local model development.