Ripple-65B: an open-weight agentic coder you can run locally

Ricursive's dense 65B tops U.S. open weights on TerminalBench 2.1, ships on HF, Together, OpenRouter and Fireworks, and cost just $32M to train.

Nowline NOV 6 4:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • U.S. open-weight SOTA on TerminalBench 2.1

    Ripple-65B Preview scores 80.5 on TerminalBench 2.1 — what Ricursive calls U.S. open-weight state of the art, edging Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4). The weights are already live on Hugging Face, with hosted endpoints on Together, OpenRouter, Fireworks and GMI Cloud, so you can point a coding agent at it today.

  • Dense, not MoE — it fits on a single DGX Spark

    It's a modernized dense Transformer (d_model 5120, 80 layers) with up to 1M-token context, deliberately avoiding Mixture-of-Experts so it loads and runs on local hardware like NVIDIA's DGX Spark. For frontier-tier open coding weights, that means no 8-GPU node just to serve it.

  • Frontier for $32M — a tenth of the last open SOTA

    Ricursive says Ripple-65B cost $32M to build from scratch, roughly one-tenth the previous open-weight agentic-coding SOTA, using ~1.4M RL rollouts (about 70x fewer than Reflection Beam) and only synthetic post-training data. The takeaway for builders: competitive open coding weights are getting far cheaper to produce — expect the release cadence to keep climbing.