Ricursive Releases Ripple-65B Preview, an Open 65-Billion Parameter Model Built for Agentic Code

Ricursive Intelligence, an American lab that describes itself as “full-stack” from model to silicon, released the weights for Ripple-65B Preview on November 4, 2026. It is a dense 65-billion parameter model, trained from scratch and released under an Apache 2.0 license, one of the most permissive available. Its main selling point: matching or outperforming American open models seven to fifteen times its size in agentic coding.

What was released

On Hugging Face, Ricursive is providing three variants: the instruct model in BF16, a quantized version in NVFP4 (the 4-bit format for recent NVIDIA GPUs), and a base model derived strictly from pre-training, without synthetic question-answering data. An auxiliary model is also included to speed up generation through speculative decoding. The model is also available via API on Together AI, GMI Cloud, and OpenRouter.

The model card details pre-training on 12 trillion tokens, primarily English and code, followed by multiple adaptation stages (long context, supervised fine-tuning, reinforcement learning) that extend the context window to one million tokens. The architecture is “thin and deep”: 80 layers, with more attention heads than usual and an increased proportion of sliding window attention layers. The model is text-only.

Ricursive emphasizes a counter-trend design choice: a dense model rather than a Mixture-of-Experts (MoE), the dominant architecture for large open models. According to the company, this choice simplified the infrastructure to the point that a team that grew from one to five people managed the training without custom compute kernels. The starting point was the open recipe from Olmo 3, developed by the Allen Institute for AI.

The claimed figures

The headline figure is a score of 80.5 on Terminal-Bench 2.1, a benchmark evaluating programming tasks conducted in a terminal. Ricursive compares it to other American open models: 80.1 for Reflection Beam (501 billion parameters), 76.4 for Meta Muse Spark Open (816 billion), 63.8 for Thinking Machines Inkling (975 billion), and 56.4 for NVIDIA Nemotron 3 Ultra (550 billion). The full picture is more nuanced: Ripple is outmatched on SWE-bench Verified (76.5 versus 81.6 for Beam), MCP Atlas, SciCode, and scientific questions on GPQA Diamond. The company itself acknowledges that reproducing these agentic evaluations is difficult, and part of its competitors’ scores were measured using its own tools.

Another significant announcement: Ricursive asserts that it developed and trained the model for $32 million, all inclusive, from refining the recipe to final training. It estimates this is ten times less than the previous top open model in agentic coding, although this comparison is based on reported compute commitments for Reflection rather than figures published by the latter.

Who is it for?

At 65 billion parameters, Ripple remains heavy, yet it is significantly more manageable than models featuring several hundred billion parameters. Ricursive notes that it can fit on a desktop workstation like NVIDIA’s DGX Spark, and provides the NVFP4 version to encourage local deployment. The company also claims 250 tokens per second on two B200 GPUs prior to quantization. For Ricursive, this model serves as a showcase: an internal variant powers its AI-assisted chip-design platform.