Ricursive releases Ripple-65B Preview, a dense Apache 2.0 coding model
A 65B dense Transformer rather than Mixture-of-Experts, pretrained from scratch on 12 trillion tokens. Ricursive reports 80.52 on Terminal-Bench 2.1, above several far larger US open models in its own table. MarkTechPost notes weak spots, including SciCode at 39.4, and quotes a full R&D cost of $32M from the company.
Why it matters: If its results hold up, it shows that a much smaller, simpler model can compete on agentic coding under a permissive license.
- Parameters
- 65B dense
- Active per token
- 65B
- Weights size
- 129.6 GB (BF16)
- Context
- Up to 1M
- License
- Apache 2.0
| Benchmark | Score | Type | Measured or published by |
|---|---|---|---|
| Terminal-Bench 2.1 | 80.52 | Self-reported | Ricursive |
| SciCode | 39.4 | Cited | Ricursive, via MarkTechPost |
Sources: huggingface.comarktechpost.com