65B Dense Model Beats 816B on Terminal Coding Leaderboard, Training Cost $32M

Ricursive Intelligence has released Ripple-65B Preview. It has 65B parameters, is a dense Transformer, was pretrained from scratch, and targets agentic coding. The company says it achieves SOTA on TerminalBench 2.1 among US open-weight models, with a score of 80.52.

Two pieces of background first. Open-weight means the weights are downloadable; training data, code, and licensing are not necessarily all open. TerminalBench 2.1 tests whether a model can complete multi-step tasks in a terminal environment—roughly having the model modify code, run commands, and debug errors on its own. US open-weight SOTA is limited to the scope of open-weight models released by US teams.

The official comparison is straightforward:

Ripple-65B Preview's lead over Beam is only 0.42 points, a small gap. The leaderboard represents only this task and cannot be directly extrapolated to general chat, math, or long-context capabilities.

More interesting is the cost. The company says the full R&D cost of Ripple-65B Preview was $32 million, starting from scratch, including the pretraining ladder, data mixture, evaluation, debugging, synthetic data, post-training recipe, and the final formal pretraining and post-training. By their estimate, this is about one-tenth of the previous open-weight SOTA for agentic coding. The model was pretrained on 12 trillion tokens and supports up to 1 million tokens of context.

Post-training did not take a brute-force route either. The company says it used only 1.4 million rollouts to climb to 80.5 points, about one-seventieth the number of rollouts of Reflection Beam. They did not buy post-training data; they built their own synthetic data pipeline, quality filtering, and coverage metrics. The RL recipe used truncated importance sampling, score centering, and a linear length penalty, together with high-quality long-horizon synthetic environments. Task difficulty increases as model capability improves.

The architecture choice also goes against the grain. Other labs use MoE extensively; Ripple used a dense Transformer. The official explanation is that this reduces the complexity of pretraining and post-training infrastructure while retaining reasoning and agent capabilities. The model has a narrow and deep structure, d_model 5120, 80 layers, more attention heads, a higher ratio of SWA to global layers, and a smaller SWA window to offset costs. It also adds FFN activation clipping, gated attention, global NoPE, and value residuals. During pretraining and post-training, MFU stays at 40%, using only simple sharding, with no custom kernels and no quantization.

The pretraining recipe came from a series of model ladder experiments. The starting point was the Olmo 3 ladder; v1.0 used StepFun's batch size and learning-rate scaling laws; v1.1 switched to a new ladder that smoothly scaled dimensions and depth; v1.2 switched from post-norm to pre+post norm; v1.3 added sigmoid gated attention; v1.4's 2-stream mHC showed no improvement at 4B scale and was abandoned; v1.5 used global NoPE and cooled the final learning rate to zero; v1.6 added value residuals. The final version switched to custom pretraining data and heavily upsampled code. The company says that at 4B model scale, this series of changes improved the compute efficiency of coding ability by up to 16x.

Availability channels are already open: Hugging Face, Together AI, GMI Cloud, and OpenRouter, with Fireworks launching later that day. The company also says the smaller parameter package can run efficiently on local devices such as NVIDIA DGX Spark.

What is really worth watching in this news may lie beyond the number 65B: the cost structure and engineering path. Dense architecture, hardware-software co-design, small-scale ladder trial and error, and RL sample efficiency. In the large-model race, 'bigger' still sells, but this leaderboard at least pours a little cold water on the parameter-scale narrative. The judgment to keep is that the 0.42-point lead and the company's self-testing still need third-party reproduction, and open weights do not equal fully open source.

Sources: Ricursive Intelligence's X post and the official Ripple-65B Preview blog. The leaderboard, cost, and rollout figures are all officially announced or estimated. Official blog: https://www.ricursive.com/ripple