Ricursive releases open coding model, claiming a lead over selected U.S. rivals

Ricursive Intelligence released Ripple-65B Preview on November 4, 2026, giving developers downloadable access to a coding-capable model it says beats selected, much larger U.S. open models on two tests. The 65-billion-parameter release carries an Apache 2.0 license, but its claimed lead comes with conflicting comparison figures and no independent confirmation of the results.

The model grew out of Ricursive’s semiconductor work. The company describes Ripple as a general-purpose variant of the proprietary model behind its AI chip-design platform—not the complete chip-design system. Its coding scores therefore should not be read as evidence that the separate platform works.

Two strong scores, not a coding crown

Ricursive reports 80.5 on Terminal-Bench 2.1, which tests work in a terminal environment, and 79.1 on SWE-bench Pro v2.0 Hard, a difficult software-engineering test. Its launch comparison puts Ripple ahead of selected U.S. open models on both. On Terminal-Bench, however, its reported advantage over Reflection’s Beam is just 0.4 percentage points: 80.5 versus 80.1.

Other results cut against a sweeping lead. Ricursive’s own table puts Meta’s Muse Spark Open ahead on DeepSWE, automation and adversarial safety tests. Meta and Reflection also score higher on the displayed math test. Reflection’s separate comparison lists several Chinese-developed models above Ripple’s reported Terminal-Bench score, although those vendor evaluations are not a controlled head-to-head test.

VentureBeat also identified inconsistencies within Ricursive’s materials. The model card gives Meta a SWE-bench score of 52.3, while the launch graphic gives it 72.4. The card combines published results with internal evaluations and acknowledges difficulties reproducing agent benchmarks. Scores also changed with the evaluation framework—the surrounding software that lets a model use tools and manage a task.

Smaller overall does not mean less work per token

Ripple uses a dense architecture: all 65 billion parameters, the learned numerical values inside the model, participate during generation. Ricursive says this simplified training compared with a mixture-of-experts design, which uses selected parts of a larger model. Beam illustrates the distinction: it has 501 billion parameters overall but activates 23 billion, versus Ripple’s 65 billion. Total size and computation per generated token are different measures.

Ricursive says it trained Ripple from scratch on 12 trillion tokens, units of text. It also reports reaching its Terminal-Bench result after 1.4 million reinforcement-learning rollouts—trial sequences used as training feedback. Reflection reported more than 100 million rollouts, but differing sequence lengths, hardware needs and evaluation expenses prevent those counts from establishing an equivalent cost saving.

$32 millionCompany-reported total

Ricursive says this includes experiments, data preparation, evaluations, troubleshooting, synthetic data, later-stage training development and final runs—not just the final training run.

Downloadable weights, a shorter hosted window

The weights are available on Hugging Face under Apache 2.0, allowing commercial use, modification and redistribution subject to the license’s conditions. There is no revenue or monthly-user threshold requiring a separate commercial agreement. Developers redistributing the model must still preserve applicable notices, provide the license and identify changes to modified files.

Download options

• BF16 instruction-tuned weights.
• Compressed NVFP4 instruction-tuned weights.
• A BF16 base model.

For hosted access, OpenRouter pricing lists $0.20 per million input tokens and $0.60 per million output tokens through GMI Cloud. That endpoint currently lists a 16,384-token context window—the amount of text it can work with at once. Developers should not mistake that service limit for Ricursive’s advertised million-token model capacity, or assume the listed price includes access to the larger window.

Local deployment needs its own checks. Ricursive names DGX Spark and MacBooks, but does not establish a specific laptop memory configuration, speed or usable context length. The base, instruction-tuned and compressed versions are distinct download choices; their memory use and performance need to be measured for the intended setup.