On Terminal-Bench, a slight lead claimed for Ripple-65B

Released as open weights under Apache 2.0, Ripple-65B Preview claims strong agentic coding scores with 65 billion parameters and $32 million in R&D.

Sixty-five billion parameters would be enough to rival several much larger American models on certain agentic programming tasks. That is Ricursive Intelligence’s claim with Ripple-65B Preview, a dense model pretrained from scratch on 12 trillion tokens. The lab releases its parameters under the Apache 2.0 license and announces a context window of up to one million tokens.

The comparison should be framed carefully: the results presented come from the lab, concern a selection of American open-weight models, and vary depending on the tests. The model distributed on Hugging Face is a general-purpose variant of a proprietary system used by Ricursive for its chip-design platform. Agentic programming is the main focus of this Preview version. A 0.4-point lead over Terminal-Bench

The headline score reaches 80.5% on Terminal-Bench 2.1, compared with 80.1% for Reflection Beam, 76.4% for Meta’s Muse Spark Open, and 56.4% for NVIDIA’s Nemotron 3 Ultra, according to Ricursive’s published table. The lead over Beam is therefore 0.4 points. It puts Ripple at the top of this specific comparison, without establishing general superiority over programming agents.

The other benchmarks paint a more mixed picture. On SWE-bench Pro v2.0 Hard, the lab reports 79.1%, ahead of the 76.5% attributed to Beam. On DeepSWE 1.1, the 44.3% result remains below Muse Spark Open’s 60%. On SWE-bench Verified, Ripple’s 76.5% is behind Beam’s 81.6%. Test versions matter: the scores from the different SWE-bench editions describe distinct evaluations.

Run conditions also come into play. The sheet indicates BF16 evaluations, with maximum effort settings, and mixes scores published by competing labs with measurements performed in-house to fill in missing data. The team acknowledges that the software driving the agent, the length of the context window, and its summarization method can shift the results. It publishes a second Terminal-Bench comparison with Terminus 2, where the scores of several competitors differ from their stated figures. Fewer stored parameters, more active parameters

The architectural choice explains part of the model’s distinctiveness. A dense Transformer here uses all 65 billion of its parameters, whereas mixture-of-experts, or MoE, models activate only a fraction for each token. In Ricursive’s table, Beam has 501 billion parameters in total but 23 billion active; Muse Spark Open has 816 billion total and 42 billion active.

Presenting Ripple as much smaller is therefore relevant in terms of total parameter count, without being enough to conclude that it requires proportionally less compute at each step. The distinction matters when assessing the resources needed to run it. Size comparisons must keep these two measures in view, rather than turning the ratio between total volumes into an automatic speed advantage.

The architecture uses 80 layers and an internal width of 5,120. The lab says it favored a dense design to simplify training and deployment infrastructure on NVIDIA hardware. Pretraining uses a text corpus made up mainly of English and code, with sequences of 8,000 tokens. Context extension then comes later in the additional training phases. The quantized format clarifies the promise of local execution

The release includes an instruction-tuned BF16 version, a quantized variant, and a BF16 base model. The main repository is about 130 GB, versus 57.2 GB for the NVFP4 repository. This second version is presented as experimental. Its quantization keeps the attention layers in BF16 and applies the NVFP4 format to the MLP layers.

These sizes provide a reference point for choosing a local machine. Memory needed for execution and context handling also has to be accounted for, in addition to the parameters. The lab cites the DGX Spark as a possible piece of hardware. The 250 tokens per second mentioned in the sheet were measured on two B200 accelerators with a batch size of eight: this measurement concerns a datacenter configuration.

An OpenRouter access option is also listed. At the time of consultation, the page shows a context window of 1,048,576 tokens, with pricing of $0.20 per million input tokens and $0.60 per output token. For download, the Apache 2.0 license file allows use, modification, and redistribution under its terms, including for commercial use. The $32 million covers full development

The budget announced in the lab presentation covers preparatory experiments, data trials, evaluations, synthetic data generation, and final training runs. The $32 million therefore represents a stated R&D cost. The claimed ten-to-one ratio versus the previous leader is based on an estimate of the competitor’s budget, without any published common accounting.

For reinforcement learning, the team says it reached 80.5% on Terminal-Bench 2.1 after 1.4 million training trajectories, or roughly seventy times fewer than Beam according to its comparison. The number of trajectories reflects the volume of experiments run, but their length and execution cost also factor into the compute spend.

Pretraining gains were pursued through sequences of increasingly large models, starting from an Olmo 3 baseline. The announced sixteenfold efficiency factor on code corresponds to a measurement at the four-billion-parameter scale. Not all tested changes were retained: a two-stream variant using mHC, which was beneficial at smaller scale, was abandoned because it produced no gain at this size.