Ricursive Intelligence released Ripple-65B Preview on November 6, an open-weight model aimed at agentic coding. The weights are available on Hugging Face under the Apache 2.0 license, and the model is also served by Together AI, GMI Cloud, OpenRouter and Fireworks.
Ripple is a 65-billion-parameter dense Transformer with 80 layers and a 5,120-dimension hidden size, pretrained from scratch on 12 trillion tokens with support for contexts up to 1,048,576 tokens. Ricursive describes it as a general-purpose variant of the proprietary model behind its AI chip design platform. Companion releases include an NVFP4 quantized checkpoint and a BF16 base model.
The company reports a self-measured 80.5 on Terminal-Bench 2.1, ahead of the figures it lists for Reflection Beam (80.1), Meta Muse Spark Open (76.4), Thinking Machines Inkling (63.8) and NVIDIA Nemotron 3 Ultra (56.4), all of which are far larger. The model card also lists 76.5 on SWE-bench Verified, 61.42 on SWE-bench Pro v1.0, 85.2 on GPQA Diamond and 98.3 on AIME 2025.
Ricursive puts total R&D cost at $32 million, which it estimates is one-tenth the cost of the previous open-weight leader in agentic coding. It says post-training used 1.4 million RL rollouts, roughly 70 times fewer than Reflection Beam, with no purchased post-training data, and that it sustained about 40% MFU without custom kernels. The card reports 250 tokens per second on two B200 GPUs at batch size 8.
Because the model is dense rather than mixture-of-experts, Ricursive says it can run on local hardware such as NVIDIA DGX Spark. For teams that want a permissively licensed coding agent they can host themselves, Ripple offers frontier-adjacent Terminal-Bench results at a size that fits on a single workstation-class system.