65B Model Tops U.S. Open-Weight Coding Leaderboard, R&D Bill $32 Million

Ricursive Intelligence has released Ripple-65B Preview. A 65B dense model, pretrained from scratch on 12 trillion tokens, supporting up to 1M context. The official positioning is straightforward: the new SOTA for agentic coding among U.S. open-weight models.

On TerminalBench 2.1, Ripple-65B Preview scores 80.52. In the comparison list, Reflection Beam 501B is 80.1, Meta Muse Spark Open 816B is 76.4, Thinking Machines Inkling 975B is 63.8, NVIDIA Nemotron 3 Ultra 550B is 56.4. Parameter counts keep getting larger, but scores do not grow with them. On this U.S. open-weight model leaderboard, parameter scale is no longer the only narrative.

The cost part is more worth looking at. Officially, the full R&D cost of Ripple-65B Preview is $32 million, including pretraining ladder development, data mixing, evaluation, debugging, synthetic data generation, post-training recipe development, and final pretraining and post-training runs. According to the official estimate, this is one-tenth the cost of the previous open-weight agentic coding SOTA. In 2023 Sam Altman said GPT-4 training cost exceeded $100 million; in 2024 Dario Amodei said frontier model training costs could reach $1 billion and predicted that in 2026–2027 pretraining alone would reach $10–100 billion. Against this backdrop, $32 million is not a small number, but it stands out sharply in the frontier-model narrative.

The architecture choice is also somewhat counterintuitive. Ricursive chose a dense Transformer, not MoE. The official rationale is to reduce pretraining and post-training infrastructure complexity while retaining reasoning and agent capabilities. The model is thin-and-deep, d_model 5120, 80 layers, with a relatively high number of attention heads, while increasing the ratio of SWA to global layers and shrinking the SWA window to offset cost. It also adds FFN activation clipping, gated attention, global NoPE, and value residuals. During pretraining and post-training, MFU stays at 40%, using simple sharding, with no custom kernels and no quantization. The small parameter envelope lets it run on local devices such as NVIDIA DGX Spark, which have high FLOPs and low memory bandwidth relative to data-center accelerators.

The RL recipe does not rely on brute force. Officially, no post-training data was purchased; they built their own synthetic data pipeline, quality filtering, and coverage metrics. The hybrid RL recipe uses truncated importance sampling, score centering, and a linear length penalty, paired with high-quality long-horizon synthetic environments. RL tasks are adjusted to model capability; as the model gets stronger, harder tasks continue to be added. In the end, climbing up from a strong SFT checkpoint took only 1.4 million rollouts to reach TerminalBench 2.1 80.5. Compared with Reflection Beam, the official claim is about 70x fewer rollouts.

The pretraining recipe came from model ladder experiments. The starting point was Olmo 3's model ladder. v1.0 implemented StepFun's batch size and learning rate scaling law; v1.1 switched to a new ladder that smoothly scales architecture dimensions and model depth; v1.2 moved from a post-norm baseline to pre+post norm, neutral at small scale and improving from 4B; v1.3 added sigmoid gated attention; v1.4's 2-stream mHC was effective at small scale but did not improve at 4B and was rejected; v1.5 switched to global NoPE and added final LR cooldown to zero; v1.6 added value residuals. The final version switched to a custom pretraining dataset with heavy upsampling of code. Measured by dolma_stack cross-entropy, these changes improved code compute efficiency by up to 16x at 4B scale. In other words, to reach code performance similar to the Olmo 3 ladder, it uses 16x less compute. The official post also notes that the Olmo 3 ladder baseline slope is flatter, implying that as scale continues to grow, the efficiency gain may become more pronounced.

Ripple-65B Preview is the general-purpose variant. An internal variant runs in Ricursive's AI chip design platform. Officially, the goal is to begin closing the loop between models and hardware. The model is already available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks later. Official website: https://www.ricursive.com/ripple. Hugging Face: http://huggingface.co/ricursiveAI/Ripple-65B-Preview. Together AI: https://www.together.ai/models/ripple-65b-preview. OpenRouter: https://openrouter.ai/ricursiveai/ripple-65b-preview.

On the open-source ecosystem side, Ricursive thanked AI2's Olmo team, Prime Intellect, RadixArk, and NVIDIA. This acknowledgment is not just courtesy. A 65B model being able to beat larger models on a public leaderboard relies on a combination of data, architecture, RL recipe, and engineering efficiency. After the model card and paper details are public, the community will probably first try to reproduce the two numbers: 1.4 million rollouts and 16x compute efficiency. They say more about the matter than parameter count.