Dense 65B models challenge MoE on agentic coding

Ricursive Intelligence released Ripple-65B Preview, reporting a top score of 80.52 on TerminalBench 2.1 for US open-weight models. Edging past Reflection Beam 501B and Meta's 816B Muse Spark Open is a notable milestone, but the architectural choice matters more than the benchmark. Ricursive opted for a dense 65-billion parameter transformer, passing over the sparse mixture-of-experts designs that dominate recent leaderboard entries.

For teams running coding agents on internal infrastructure, dense models at this scale solve practical hardware headaches. Sparse mixture-of-experts models reduce active computation per token, but keeping hundreds of billions of idle weights in memory requires complex multi-node clusters and high-bandwidth interconnects. A dense 65B model fits into simpler hardware configurations, making local terminal execution loops faster and much cheaper to host on-premise.

Ricursive attributes its reported $32 million training cost to small-scale model laddering and synthetic reinforcement learning environments, citing 70 times fewer rollouts than Reflection Beam. Launch benchmarks always need verification, but showing that a compact dense model can match mammoth architectures on agentic tasks keeps self-hosted coding agents practical.