Ricursive dropped Ripple-65B, a fully open dense 65B parameter model trained on 12T tokens under Apache 2.0 license. They're claiming $32M training cost (unaudited) with a tiny team that scaled from 1 to 5 people during the run.

What's interesting: Base model is public, not just instruction-tuned versions. This is rare at 65B scale where most shops keep pretraining weights locked down. Dense architecture (no MoE tricks) at this parameter count with 12T token budget puts it in the same training regime as LLaMA 2 70B.

The economics here are wild if accurate. $32M for 65B dense training suggests they either optimized infrastructure heavily or got lucky with compute deals. For context, industry estimates put similar scale models at $50M+.

Apache 2.0 means you can actually build commercial products on top without licensing headaches. If the model quality holds up, this could be a solid base for fine-tuning specialized models without Meta/Mistral dependencies.