Ricursive Intelligence has released Ripple-65B Preview, a general-purpose open-weight model derived from work on the company’s AI chip-design platform. The November 4 release gives developers another substantial coding model to evaluate, with downloadable checkpoints and access through multiple inference providers.
The launch announcement describes a dense 65-billion-parameter Transformer trained from scratch on 12 trillion tokens. Ricursive chose a dense architecture rather than a mixture of experts and says its hardware and model design were developed together. The company reports an 80.52 result on TerminalBench 2.1 and a total research-and-development cost of $32 million. Those are Ricursive’s reported figures, rather than results independently reproduced for this article.
Weights, formats and context
The official model card lists Apache 2.0 licensing and describes instruction-tuned BF16 and NVFP4 checkpoints alongside a BF16 base model. It specifies a context window of up to 1,048,576 tokens. The card also distinguishes the final instruction model from the earlier pretraining checkpoint, a useful detail for teams choosing which artifact to download.
Ricursive acknowledges that agentic benchmark results change with the execution harness and configuration. Its comparison includes both published scores and evaluations run internally to fill gaps. That makes the release interesting evidence about a particular system configuration, while leaving room for developers to reach different results in their own environments.
What developers should measure
Our practical reading is that the most useful comparison begins with an entire task. A model that generates tokens quickly can still take longer overall if it makes extra terminal calls, repeatedly edits the wrong file or needs several attempts to finish a change. Evaluation should record completion rate, elapsed time, inference cost and the amount of human repair required.
The same applies to the large context window. A repository fitting into a prompt does not establish that the model will find every relevant dependency. A realistic trial should include a change that crosses modules, requires a test update and exposes an error that cannot be fixed by editing the first file encountered. Keep the repository revision and execution environment fixed so the comparison remains useful.
Open weights also let infrastructure teams examine a different tradeoff: whether predictable local capacity is more valuable than a managed endpoint. That calculation includes memory, throughput under concurrency, operational support and upgrade work. Ripple’s release provides artifacts for that investigation; the result will depend on the workload and hardware available.
For buyers, the immediate development is a new deployable option with a clear license and documented checkpoints. For researchers, the accompanying account of architecture and training offers a starting point for investigating how a smaller dense model can compete on demanding coding tasks.