65B dense model takes U.S. open-weight agentic coding SOTA at a cost of $32 million

Ricursive Intelligence has released Ripple-65B Preview, a 65B dense model pretrained from scratch. The release page makes its main points clear: a new SOTA for U.S. open-weight models on agentic coding tasks, and a detailed, line-item training cost accounting.

First, the scores. Ripple-65B Preview scores 80.52 on TerminalBench 2.1, beating several larger U.S. open-weight models: Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, Thinking Machines Inkling 975B at 63.8, and NVIDIA Nemotron 3 Ultra 550B at 56.4. Note the qualifier: this is SOTA within the U.S. open-weight scope, not a ranking of all models worldwide. TerminalBench 2.1 measures an agent's ability to complete multi-step coding tasks in a terminal environment; a high score does not mean it is strong in every coding scenario.

Model specs: pretraining used 12 trillion tokens and supports up to 1M context. Architecturally, Ricursive chose a dense Transformer route, not MoE. Dense models activate all parameters at every inference step, which is more straightforward engineering-wise; MoE uses sparse activation to reduce compute, but training and serving are more complex. Ripple's choice is thin-and-deep: d_model is 5120, with 80 layers total, attention heads were over-provisioned, and it also increased the ratio of SWA to global attention layers and shrank the SWA window to offset cost. It also added FFN activation clipping, gated attention, global NoPE, and value residuals.

Cost is the most noteworthy part of this release. Ricursive says the total R&D cost of Ripple-65B Preview was $32 million, about one-tenth of the previous agentic coding open-weight SOTA. This number covers pretraining ladder development from scratch, data mixture, evaluation, debugging, synthetic data generation, post-training recipe development, and the final formal pretraining and post-training runs. The release page also cites Sam Altman's statement that GPT-4 training cost more than $100 million, and Dario Amodei's prediction that frontier model training costs could reach the $1 billion to $10 billion level. By comparison, the $32 million ledger is indeed rare.

Post-training did not rely on brute-force RL scaling. Ricursive built its own synthetic data pipeline, quality filtering, and coverage metrics, and did not buy post-training data. The RL recipe uses truncated importance sampling, score centering, and a linear length penalty, together with high-quality long-horizon synthetic environments. As model capability improved, task difficulty kept increasing. In the end, only 1.4 million rollouts were used to push the TerminalBench 2.1 score to 80.5, which the company says is about 70x fewer rollouts than the previous SOTA, Reflection Beam.

The pretraining recipe came from a series of model ladder experiments. The starting point was the Olmo 3 model ladder; the team tested changes at small scale, focusing on coding ability. v1.0 introduced StepFun's batch size and learning rate scaling laws, v1.1 switched to a new ladder method, v1.2 moved from post-norm to pre+post norm, v1.3 added sigmoid gated attention, and v1.4's 2-stream mHC was effective at small scale but showed no improvement at 4B scale, so it was rejected; v1.5 switched to global NoPE and cooled the final learning rate to zero, and v1.6 added value residuals. The final version switched to a custom pretraining dataset and heavily upsampled code. The company says these changes improved code compute efficiency by up to 16x at the 4B model scale; measured by dolma_stack CE, reaching code performance similar to the Olmo 3 ladder required only about 1/16 of the compute.

On deployment, Ripple's dense architecture and smaller parameter package let it run on local devices such as NVIDIA DGX Spark. Training maintained 40% MFU throughout, used simple sharding, and had no custom kernels or quantization. For developers, the model is already available on Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks coming online later that day. The release page is at Ricursive official website, where weights and trial access can be found directly.

What is worth noting about this release is that it puts parameter scale, training cost, RL sampling efficiency, and local deployment on the same table. A 65B dense model being able to squeeze into the 80-point range on TerminalBench 2.1 shows that training recipes and task design are becoming more important. Whether it can reproduce this score on a broader range of coding tasks still has to wait for developers to try it themselves.