A Palo Alto, California startup, Recursive, has previewed Ripple-65B, a 65B-parameter dense model trained from scratch, and disclosed a training cost of approximately $32 million. Our judgment is that the value of this disclosure lies not merely in the model’s scale, but in showing how companies outside the leading labs are trying to lower the barrier to large-model training through co-design across chips and training workflows.
What this is
Ripple-65B is Recursive’s preview release of a 65B-parameter dense large language model. Unlike sparse models that selectively activate parameters by task, a dense model uses all parameters for every piece of text it processes. Recursive attributes its $32 million training spend to its long-term focus on chip design. The model remains in preview, however, and has not yet received adequate independent benchmarking or validation on real-world tasks.
The post also lists models including NVIDIA’s Nemotron 3 Ultra-550B, Thinking Machines Lab’s Inkling-975B, Meta’s Muse Spark Open-816B, and Reflection AI’s Beam-501B for comparison. Notably, many of these names and parameter counts come from their publishers or community compilations. Some information still requires cross-verification and cannot be treated as a performance ranking under identical conditions.
Industry view
If the $32 million cost figure is accurate and reproducible, it reinforces a broader trend: differentiation in large-model training is shifting from “Who has the bigger model?” to how chips, compilers—the tools that translate model code into instructions executable by hardware—and training systems work together. For enterprises with predictable compute requirements and a willingness to participate in early testing, this could offer a new way to reduce the cost of procuring models or building them in-house.
The risks are equally clear. First, training cost does not equal deployment cost, much less the total cost of obtaining usable capability within an enterprise. Inference, data governance, integration, and ongoing maintenance may account for larger expenditures. Second, the combination of a $32 million budget and 65 billion parameters needs to be validated through data quality, training efficiency, and benchmarking. The post itself acknowledges the risk of benchmark cherry-picking, so it is better viewed as a signal of a technical direction than as a performance conclusion.
Impact on regular people
For enterprise IT: Falling model-training costs may make it easier for mid-sized companies to evaluate specialized models. Selection should still prioritize stability, deployment cost, and performance on business-critical tasks rather than parameter count.
For working professionals: Cheaper inference is expected to continue lowering the barrier to using professional tools, but Ripple-65B remains in preview and will not directly reshape workflows in common office software in the short term.
For the consumer market: If the cost advantage holds, local and customized models may become more common, bringing broader availability of privacy-preserving and offline use. That outcome will still depend on endpoint compute, licensing models, and service quality.