Ricursive Intelligence, an artificial intelligence laboratory working on chips and models together, has released the Ripple-65B Preview, an open-weights large language model with 65 billion parameters. According to the company, with a score of 80.52 on the TerminalBench 2.1 benchmark, it is the best result among American open models in agentic coding—that is, in tasks where a software agent works alone on the terminal.
The score surpasses larger models, such as Reflection Beam with 501 billion parameters (80.1), Meta Muse Spark Open with 816 billion (76.4), Thinking Machines Inkling with 975 billion (63.8), and NVIDIA Nemotron 3 Ultra with 550 billion (56.4). These values are stated by Ricursive and have not been independently verified.
A dense, not expert, architecture

Image: datasets — https://huggingface.co/datasets/pratikgx/ripple-65b-preview-m5-pro-mac-mini-bench
Ripple is a dense model: all 65 billion of its parameters work for every token, unlike the Mixture-of-Experts (MoE) models used by many labs. Ricursive chose this path to simplify training and serving infrastructure. The model is pre-trained on 12 billion tokens and handles up to one million context tokens.
The structure is narrow and deep: the hidden dimension is 5,120 and there are 80 layers. To improve reasoning, the company increased the number of attention heads and modified the ratio between sliding window attention and global attention. According to Ricursive, the model maintains a 40% compute utilization efficiency during training and post-training, without custom kernels or quantization.
The compact size allows the model to run on local devices such as the NVIDIA DGX Spark, the machine Ricursive cites as the benchmark for local use.
Costs and method
Ricursive states that it spent $32 million to develop the model, including research and development, testing on smaller models, synthetic data generation, and final training. According to the company, this is one-tenth the cost of the previous state-of-the-art in agentic coding, an estimate made from memory by the lab. This figure is also significant for comparison with Dario Amodei's forecasts, which estimated costs of up to $10-100 billion by 2026-2027 for pre-training frontier models alone.
Post-training was built without purchasing data: Ricursive created its own synthetic data pipelines, quality filters, and coverage metrics. The reinforcement learning mechanism achieved a score of 80.5 with 1.4 million releases, about seventy times fewer than Reflection Beam, according to the company.

Image: ricursiveAI — https://huggingface.co/ricursiveAI/Ripple-65B-Preview
The model was built on open-source foundations, including the work of the Olmo team from the Allen Institute for AI, Prime Intellect, RadixArk, and NVIDIA.
Where it can be used
Ripple-65B Preview is available under the Apache 2.0 license on Hugging Face, in an instruction-tuned version (BF16 and NVFP4) and as a base model. The model is also hosted on Together AI, OpenRouter, GMI Cloud, and Fireworks. No price is indicated in the consulted sources for APIs.
Ricursive was founded as a lab that builds systems capable of improving themselves and has already been presented on AlexTech as a company aiming to design chips using artificial intelligence. Ripple is the company's first public model, while the internal version is what powers its chip design platform.
The stated performance refers to a specific benchmark and does not measure all activities. The self-reported score should be read as an indication of the company's work, pending independent verification.

No comments yet. Be the first!