Ricursive Intelligence Releases Ripple-65B-Preview Dense Model
Ricursive Intelligence has launched Ripple-65B-Preview, a new 65-billion-parameter dense open-weights model featuring a thin and deep architecture. The preview release has undergone light post-training with a primary focus on agentic coding capabilities. This release fills a long-awaited gap in the open-source community for a new ~70B dense model that can compete with much larger models. It provides researchers and developers with a powerful, versatile tool for local deployment and further customization. The model utilizes a "thin & deep" aspect ratio of 80 layers by 5120 dimensions, a 4:1 sliding window to global attention ratio, FFN activation clipping, gated attention, global NoPE, and value residuals. GGUF formats are available via AtomicChat, along with a custom llama.cpp patch for local execution.
## BACKGROUND
Dense models store and activate all of their parameters for every token generated, unlike mixture-of-experts (MoE) models which route tokens through a subset of parameters. Open-weights models allow community users to inspect, run, and fine-tune the model parameters locally on their own hardware.