Ripple-65B Preview focuses on an open model tailored for code

Ricursive Intelligence unveiled Ripple-65B Preview, an open language model with 65 billion parameters. The company aims for agentic programming, with a total development cost of 32 million dollars.

Ricursive Intelligence, a company specializing in AI chip design, has presented Ripple-65B Preview, an open language model trained on 12 trillion tokens and released under the Apache 2.0 license. The company claims it scores 80.52 on TerminalBench 2.1, the highest score among open American models, a result that mainly matters for agent-assisted coding uses. In short, Ripple aims to secure a place in developers' toolkits, not just in demos that shine at conferences.

Un modèle dense, pas une usine à experts

Ripple-65B Preview is based on a dense Transformer architecture with 65 billion parameters, whereas many recent models rely on a mixture-of-experts approach, meaning an architecture that activates only part of the network depending on the task. Ricursive says it chose this path to simplify deployment, especially on personal hardware, even if the model remains optimized for Nvidia GPUs. This positioning may influence companies wanting to reduce operational complexity without giving up a large model.

The company also states that development and training cost a total of 32 million dollars, including research and development. This figure places Ripple in a more accessible range than some giant market programs, without implying the model is cheap to operate. Promises of efficiency are sometimes hard to keep, but the calculation remains very concrete for infrastructure teams.

Des résultats solides, mais pas les meilleurs partout

On benchmarks published by Ricursive, Ripple-65B Preview stands out on TerminalBench 2.1 but lags behind Inkling 2 Small on other tests. The provided comparison shows 79.1% versus 86.3% on SWE-Bench Pro v2.0 Hard, 44.3% versus 72.1% on DeepSWE v1.1, and 71.67% versus 85.4% or 85.0% on MCP Atlas. In other words, the model does not dominate the entire board but shows a consistent profile for certain agentic coding tasks.

The available sources do not specify exactly the comparison protocol between Ripple and Inkling, nor the test execution conditions [to be verified]. They do indicate that Ripple partially uses Terminus 2 and partially OpenCode as harness, meaning the evaluation tooling around the model. This detail matters because a benchmark is only valuable if one knows how it was conducted.

Pourquoi cette sortie peut intéresser le marché du code

With an Apache 2.0 license and availability on Hugging Face, Ripple-65B Preview targets an audience wanting to test, adapt, and possibly host the model themselves. This type of distribution attracts product teams, researchers, and integrators seeking an open alternative to closed models, especially for software development uses. The real issue is therefore not only raw performance but deployment freedom and integration cost.

Ricursive does not just say it built a model; the company also tries to show that a chip-specialized player can advance on the software front with a credible proposal. It remains to be seen if the benchmarks translate into real gains in production environments, where prompts are sometimes less polished than in slides.

By the numbers

  • 12 trillion tokens — pretraining of Ripple-65B Preview, according to Ricursive.
  • 65 billion parameters — dense architecture of the model.
  • 32 million dollars — total development and training cost, according to the company.
  • 80.52 — announced score on TerminalBench 2.1.
  • 2026 — Preview release date, according to the source.

To read