Ricursive Intelligence has released Ripple-65B Preview, a dense 65-billion-parameter model aimed partly at agentic coding. The November 4 release includes instruction-tuned weights in multiple formats and a separate base checkpoint. This report was published November 8.
What the model card says
Ricursive describes 12 trillion pretraining tokens and support for up to 1,048,576 tokens of context. It reports 80.5 on Terminal-Bench 2.1 and provides details of the evaluation harness. The team also warns that benchmark results can shift with the harness and settings. These are the developer's measurements, not an AiBrainWorX test or a guarantee of a lead in other workflows.
The listed model license is Apache 2.0. The BF16 repository lists around 130 GB of files; the available lower-precision version is a separate artifact with its own requirements. A long advertised context does not remove the memory and latency cost of using it.
For local builders
This is an interesting inspectable coding model, but it is not a drop-in fit for every consumer PC. Confirm the exact quantization, runtime support, full memory use and license before downloading. On a 12 GB GPU, assume a substantial offload or different hardware will be needed rather than expecting an in-VRAM fit. We have not installed or timed it.
Original reporting; claims attributed to their sources. No local tests implied.
