Ricursive releases Ripple preview weights for coding agents
Ricursive Intelligence has released Ripple-65B Preview under Apache 2.0, allowing developers to download and adapt its model weights. The dense model focuses on agentic coding while retaining a general-purpose design. Base and instruction-tuned checkpoints offer different starting points for further work. The release also includes quantised weights, an auxiliary generation model and hosted access through several providers.
Artificial Intelligence··Night
Preview weights become downloadable
Ricursive Intelligence, a lab developing models and chip-design systems, released the instruction-tuned Ripple-65B Preview weights on November four. The model’s Apache 2.0 licence allows developers to download and adapt the distribution. Ripple uses a dense Transformer with sixty-five billion parameters, rather than activating only a small subset through a mixture-of-experts design. Its instruction-tuned version is distinct from the base checkpoint from pretraining alone. The preview emphasizes agentic coding: software tasks in which a model works through a sequence of actions. The weights release makes that model available for local deployment as well as access through hosting providers.[1], [2]
Checkpoints support different deployment paths
Alongside the BF16 version, Ricursive supplies an NVFP4 quantised checkpoint and an auxiliary model for speculative decoding, which uses proposed continuations to accelerate generation. The company says pretraining used twelve trillion tokens, mainly English text and code. The architecture has eighty layers and a hidden width of 5120. Later training includes midtraining, long-context adaptation, supervised fine-tuning and reinforcement learning. Ricursive describes the release as a general-purpose version of the proprietary model used in its chip-design platform. It also characterizes the preview’s post-training as relatively light, with coding agents receiving particular attention. The training recipe starts from Olmo three, developed by the Allen Institute for Artificial Intelligence. The lab says it trained the model from scratch.[1], [2]
Local weights and hosted access coexist
The base checkpoint preserves the state reached after pretraining, while the instruct checkpoint adds later training for responding to requests. Developers can select that starting point when adapting the model themselves. Ripple handles text, and its architecture combines more attention heads with a greater proportion of sliding-window attention layers. These layers process limited stretches within the larger context. The model card documents support up to one million tokens; a hosting provider can expose its own limit. Ricursive lists Together AI, GMI Cloud and OpenRouter as hosted access options. Downloadable checkpoints offer a separate route for organizations running the model under their own deployment arrangements.[1], [2]
