News · 2026-11-05

Ricursive releases Ripple-65B Preview and reports $32 million in full development costs

Ricursive Intelligence released Ripple-65B Preview on November 4 under Apache 2.0 and said the complete development effort cost $32 million. Its launch pairs a reusable coding-model checkpoint with an unusually broad cost claim, while the headline agent-task results remain company-reported rather than independently reproduced.

Key facts

• Ripple is a dense 65-billion-parameter model released as a preview on November 4.
• Ricursive describes $32 million as its “full R&D cost starting from scratch.”
• Its headline terminal-agent result is about 80.5% task resolution, with two differently constructed comparison rows.
• The primary source is Ricursive’s release and methodology, supported by its model card.

The release offers two stories that deserve separate examination: whether the model is useful, and whether its development approach is economically repeatable. A good result on one agent benchmark would not validate an accounting comparison. An attractive development budget would not establish that the model beats alternatives across everyday work.

Ripple’s most prominent test puts an agent inside a command-line environment and checks whether it finishes practical tasks. This is closer to asking an apprentice to repair a broken workshop than asking someone to finish a line of code. The apprentice must inspect the situation, choose tools, change the environment and satisfy the final inspection. Its success depends on the agent software around the model as well as the model itself.

The Terminal-Bench maintainers revised 28 of 89 tasks in version 2.1 to address drifting dependencies, resource limits and mismatches between instructions and tests. That is not a criticism specific to Ripple. It is direct evidence that the test conditions matter when interpreting a narrow numerical lead.

Ricursive’s card presents one row assembled from self-reported results where Ripple is near-tied with Reflection Beam. Another row uses Ricursive’s Terminus 2 setup and shows a larger advantage over Beam. Those are different comparisons. The second is the closer common-harness comparison, but Ricursive still ran it. The launch materials reviewed for the dossier did not provide an inspectable public job, full trajectories or an independently accepted leaderboard result for Ripple.

The broader table is also mixed. Ripple does not win every coding or scientific task in its own card. Its advertised leadership is scoped to a particular comparison among United States open-weight releases. Removing that qualifier turns a narrow company claim into a global claim the evidence does not support. The Hacker News discussion focused heavily on that wording and on how much a sub-percentage-point margin should mean before replication.

The cost claim is more interesting than a final training bill. Ricursive explicitly includes developing small-model recipes, data mixing, synthetic examples, evaluations, debugging, post-training methods and the full-scale runs. It does not provide line-item treatment of salaries, depreciation or other accounting assumptions. Its comparison with another lab’s publicly reported compute commitment therefore cannot establish a like-for-like cost reduction.

The training account says the final model was pretrained from scratch, while the smaller-model ladder used to choose a recipe began from Olmo 3’s research approach. Those statements can coexist. Following another laboratory’s recipe-development framework is different from loading that laboratory’s trained parameters. The release acknowledges research and software contributions rather than demonstrating a model built without intellectual predecessors.

Ricursive also emphasizes a dense architecture and synthetic tasks that become harder as the model improves. In a dense model, all parameters participate in generating each token; in a mixture of experts, selected specialists participate instead. The company argues that its dense design simplified training and serving. That is a plausible engineering tradeoff, while its reported efficiency gains remain internal measurements.

Access is real but hardware requirements need restraint. The Apache license file supports broad reuse. The release offers higher- and lower-precision variants and points to specialized local hardware. It reports serving performance on two B200 accelerators, but the reviewed sources do not establish a minimum graphics-memory requirement. A verified weight-download size was unavailable in the dossier and could not be checked during synthesis, so no size is inferred from parameter count.

The local-model community thread discusses quantization, context memory and downloads without establishing an independent benchmark rerun. Those practical questions may decide adoption more than the launch ranking. Ripple is a substantive release and a testable economic claim. Its next milestone is inspectable performance on the released checkpoint under conditions others can reproduce.

Primary source, verified: read the paper →

Ground Truth. (2026, November 5). Ricursive releases Ripple-65B Preview and reports $32 million in full development costs. Ground Truth. https://groundtruth.day/news/ricursive-ripple-65b-preview-r-and-d-cost-claim.html

@misc{groundtruth:ricursive-ripple-65b-preview-r-and-d-cost-claim,
  title  = {Ricursive releases Ripple-65B Preview and reports $32 million in full development costs},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/ricursive-ripple-65b-preview-r-and-d-cost-claim.html}
}

Topics: open-weights · coding · models · ricursive · training-costs