Ricursive Intelligence · 2026-11-04 · major
Ripple-65B Preview — open-weight 65B coding model from Ricursive
In short
Ripple-65B Preview is a 65B dense open-weight model from Ricursive Intelligence, tuned for agentic coding. It scores 80.5 on Terminal-Bench 2.1 (self-reported) and ships under Apache 2.0.
Ricursive Intelligence · 2026-11-04 · 4 sources checked
Ricursive's first public model is a 65B dense, Apache-2.0 model built for agentic coding.
Key specs
| Parameters | 65B dense |
|---|---|
| Terminal bench 2.1 | 80.5 |
Quick facts
| Maker | Ricursive Intelligence |
|---|---|
| Architecture | 65B dense Transformer, 80 layers |
| License | Apache 2.0 |
| Context window | Up to 1M tokens |
| Training data | 12T tokens |
| Available on | Hugging Face, Together AI, GMI Cloud, OpenRouter |
What is it?
Ripple-65B Preview is an open-weight, instruction-tuned language model from Ricursive Intelligence, an AI lab whose proprietary sibling model powers its chip-design platform. Ricursive calls it the U.S. open-weight state of the art on agentic coding. The weights are on Hugging Face.
How does it work?
The model is a dense Transformer with 80 layers and d_model 5120, pretrained on 12T tokens of mostly English text and code. Post-training adds midtraining, long-context training, SFT and RL aimed at agentic coding, with about 1.4M RL rollouts. Context reaches 1M tokens.
Why does it matter?
Ricursive reports 80.52 on Terminal-Bench 2.1, against 80.1 for Reflection Beam 501B and 76.4 for Meta Muse Spark Open 816B, with a far smaller model. The numbers are self-reported. Hacker News commenters note the lead is small, the comparison leaves out Qwen models, and a 65B dense model is slower locally than MoE rivals.
Who is it for?
Teams that want U.S.-made open weights for coding agents
Frequently asked questions
- Is Ripple-65B Preview open source?
- Ripple-65B Preview is open-weight and its Hugging Face model card lists the Apache 2.0 license. Ricursive Intelligence publishes the instruction-tuned BF16 weights and lists related releases: a base model, an NVFP4 quantized version and a DFlash2 variant. Training data and code are not described as open.
- Where can I run Ripple-65B Preview?
- Ricursive says Ripple-65B Preview is available on Hugging Face, Together AI, GMI Cloud and OpenRouter, with Fireworks listed as going live later on launch day. The company also says it runs locally on an NVIDIA DGX Spark. As a 65B dense model it needs far more memory than small models.
- How does Ripple-65B Preview compare with other open models on coding?
- On Terminal-Bench 2.1 Ricursive reports 80.52 for Ripple-65B Preview, ahead of Reflection Beam 501B at 80.1, Meta Muse Spark Open 816B at 76.4, and NVIDIA Nemotron 3 Ultra 550B at 56.4. These are the company's own figures, and Hacker News commenters question the narrow comparison set.
- How much did Ripple-65B Preview cost to build?
- Ricursive Intelligence states that full R&D for Ripple-65B Preview cost about $32 million, which it estimates is one-tenth of the prior open-weight state of the art on agentic coding. It also says the model reached its results with about 70 times fewer RL rollouts than Reflection Beam.
Try it
huggingface.co/ricursiveAI/Ripple-65B-Preview