Ripple-65B Preview Hits US Open-Weight Coding SOTA at 80.52

Ricursive Intelligence officially launched Ripple-65B Preview on November 4, 2026, a 65-billion-parameter dense Transformer trained on 12 trillion tokens that delivers a new US open-weight state-of-the-art result of 80.52 on TerminalBench 2.1 for agentic coding. The lab achieved this milestone at a reported total R&D cost of approximately $32 million, roughly one-tenth the expense of prior open-weight coding models, while using only 1.4M reinforcement learning rollouts across its 80-layer architecture.

Release Date and Availability

Officially released on November 4, 2026, Ripple-65B Preview is now available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks deployment scheduled for later the same day. The model is a general-purpose variant of the proprietary system that powers Ricursive's internal AI chip design platform. It supports context lengths up to 1 million tokens and was designed to run efficiently on local devices such as NVIDIA DGX Spark, which feature high compute FLOPs with lower memory bandwidth relative to datacenter accelerators.

Benchmark Performance and Architecture

Ripple-65B Preview surpasses significantly larger US open-weight models on TerminalBench 2.1, including Reflection Beam 501B (80.1), Meta Muse Spark Open 816B (76.4), Thinking Machines Inkling 975B (63.8), and NVIDIA Nemotron 3 Ultra 550B (56.4). Unlike many competing labs that rely on Mixture-of-Experts (MoE) architectures, Ricursive chose a modernized dense Transformer design to reduce infrastructure complexity while maintaining strong reasoning and agentic performance.

The model uses a thin-and-deep aspect ratio with a narrow d_model of 5120 and 80 total layers. To boost reasoning capabilities, Ricursive overprovisioned attention heads while increasing the sliding-window-attention-to-global layer ratio and reducing SWA window size. The architecture incorporates FFN activation clipping, gated attention, global NoPE, and value residuals. This design maintained a solid 40 percent model FLOPs utilization throughout pretraining and posttraining, even without custom kernels or quantization.

Training Efficiency and Cost Optimization

Ricursive developed and trained Ripple-65B Preview for $32 million, a figure that includes all R&D costs from scratch: pretraining ladder development, data mixing, evaluations, debugging, synthetic data generation, posttraining recipe development, and the final hero runs. The lab estimates this represents one-tenth the cost of the previous open-weight SOTA on agentic coding. For context, OpenAI CEO Sam Altman disclosed in 2023 that GPT-4 training costs exceeded $100 million, and Anthropic's Dario Amodei projected frontier models could reach $10-100 billion in training costs by 2026 to 2027.

Rather than investing in brute-force RL scaling, Ricursive iterated carefully on supervised fine-tuning and RL recipes with a focus on compute efficiency. The team built its own synthetic data pipelines rather than purchasing posttraining data, using quality filtering and coverage metrics. The mixed RL recipe employs truncated importance sampling, score centering, and linear length penalty with high-quality long-horizon synthetic environments. RL tasks were tuned continuously as model capabilities improved, producing a steep climb from a strong initial SFT checkpoint. The final result required only 1.4 million rollouts to reach 80.5 on TerminalBench 2.1, roughly 70 times fewer rollouts than Reflection Beam.

Model Ladder Development and Optimization

Ricursive optimized Ripple's pretraining recipe through a series of model laddering experiments, starting from the Olmo 3 baseline from the Allen Institute for AI. The team tested interventions at small scale, focusing on improving coding capabilities. Key changes across ladder versions included implementing StepFun batch size and learning rate scaling laws, switching from post-norm to pre+post norm, adding sigmoid gated attention, adopting global NoPE with a final learning rate cooldown to zero, and incorporating value residuals. The final ladder also switched to a custom pretraining dataset with substantially upsampled code.

Together, these compounded changes improved compute efficiency on code by up to 16 times at the 4B model scale, as measured by dolma_stack cross-entropy. In other words, the final Ripple model ladder achieves similar code performance to the Olmo 3 baseline with 16 times less compute. The team notes that the compute efficiency improvement increases with additional scale, as evidenced by the shallower slope of the Olmo 3 baseline compared to later ladder versions.

What This Means for Open-Weight AI

Ripple-65B Preview demonstrates that competitive open-weight performance on agentic coding is achievable at a fraction of the cost and compute previously assumed necessary. By co-optimizing model architecture, training recipes, and hardware compatibility, Ricursive has set a new efficiency benchmark for labs building self-improving systems. The model's availability across multiple platforms lowers the barrier for researchers and developers working on coding agents and related applications. As Ricursive scales toward the frontier with its internal chip design platform, the open-source ecosystem stands to benefit from continued architectural and training innovations that emerge from this full-stack approach.