Ricursive AI Releases Ripple-65B Preview: A 65B Dense Open-Weight Model Scoring 80.5 on Terminal-Bench 2.1

Ricursive Intelligence released Ripple-65B Preview, a 65B dense open-weight model pretrained from scratch for agentic coding and offered under Apache 2.0. It's a general-purpose cut of the proprietary model behind Ricursive's AI chip design platform. The lead claim: 80.52 on Terminal-Bench 2.1, the top self-reported score among US open-weight models, ahead of MoE systems with 501B–975B total parameters.

Pretraining ran on 12T tokens with up to 1M context. Ricursive puts full R&D cost at $32M — ladder experiments, data mixing, evals, synthetic data, and final runs — and estimates that at roughly one-tenth the cost of the prior open-weight agentic coding leader. Three checkpoints ship: instruction-tuned BF16, experimental NVFP4, and a BF16 base.

We're taking the dense-over-MoE choice seriously. Ricursive argues dense Transformers cut infrastructure complexity: 40% MFU with simple sharding, no custom kernels, no quantization during training. The model is thin and deep — d_model 5120 across 80 layers — moving from an Olmo 3 Llama-like layout toward a Gemma 3N-like design for stability, with gated attention, global NoPE, value residuals, and FFN activation clipping.

The RL efficiency claim is the one I'll keep repeating if it holds up. Ripple hit 80.5 on Terminal-Bench 2.1 with 1.4M rollouts; Ricursive puts Beam at over 100M — roughly 70× more. If you can reach frontier-ish agentic coding with two orders of magnitude fewer rollouts, the economics of open agent models change. Post-training used midtraining, long context (up to 2M sequence length), SFT, and mixed RL with truncated importance sampling, score centering, and a linear length penalty. No purchased post-training data — synthetic pipelines only.

Eyes open on the scoreboard. Ripple leads Ricursive's own table on Terminal-Bench 2.1 and SWE-bench Pro v2.0 Hard (79.1), but trails on SWE-bench Verified (76.5 vs Beam's 80.9) and looks soft on SciCode (39.4), DeepSWE 1.1 (44.3), and MCP Atlas. A model tuned for terminal agents is not a science-coding generalist. When Ricursive reran Terminal-Bench with the official Terminus 2 harness at 128K, most rivals dropped; Muse Spark Open rose to 78.7 and still sat below Ripple's 80.5.

The NVFP4 build aimed at DGX Spark-class local hardware is the product hook. Datacenter BF16 hits about 250 tok/s on 2× B200 at batch 8 before quantization. Hosting is listed via Together AI, GMI Cloud, and OpenRouter at launch (Together marked coming soon), with Fireworks later. Weights on Hugging Face.

If a $32M dense 65B can own Terminal-Bench against half-trillion MoEs, the open-weight race may pivot from parameter theater to task-shaped pretraining and ruthless RL efficiency. The harm case is familiar: stronger agentic coding models lower the cost of automated vulnerability discovery and repo-scale change as fast as they help legitimate engineering.

Ricursive builds AI tooling for chip design; Ripple is the open general-purpose sibling of that internal model. Comparison set in their materials includes Reflection Beam, Meta Muse Spark Open, Thinking Machines Inkling, and NVIDIA Nemotron 3 Ultra — mostly MoE or hybrid architectures with far larger total parameter counts.