Ricursive Intelligence releases Ripple-65B Preview, a 65B model with 1M context
Ricursive Intelligence announced Ripple-65B Preview, a dense Transformer with 65B parameters trained on 12 trillion tokens and supporting up to 1M context length. The model is offered in BF16 and NVFP4 numerics, with an instruction‑tuned version and a base BF16 checkpoint.
Key points
- Ripple-65B Preview has 65B parameters, 12 trillion pretraining tokens, and 1M context length
- Scores 80.52 on TerminalBench 2.1, beating larger US open‑weight models
- Training efficiency improves 16x vs Olmo 3 ladder, reaching 55x at full scale
In agentic coding benchmarks the model scored 80.52 on TerminalBench 2.1, outpacing larger open‑weight models such as Reflection Beam 501B (80.1) and Meta Muse Spark Open 816B (76.4). Compute efficiency on code improved 16x versus the Olmo 3 ladder and rises to 55x for the full run, while maintaining 40% MFU throughout training. The architecture uses a narrow 5120 hidden dimension across 80 layers and runs at about 250 tok/s on 2xB200 GPUs with batch=8. The team grew from 1 to 5 people and the model can fit on consumer hardware like an NVIDIA DGX Spark, with a public NVFP4 checkpoint and colab to help users start.
ricursiveAI/Ripple-65B-Preview · Hugging Face
huggingface.co · 4 November 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Tencent releases Hy4 open-weight model, claims top-tier Chinese AI · 2 src
- Seven open specialist models under 10B parameters outperform larger systems in niche tasks · 1 src
- oMLX 0.7.4 adds prefill accelerator using a second Mac for large models · 1 src
- OpenAI’s 2022: the year ChatGPT and AI art sparked debates over usage and ethics · 1 src
- Axion1-0/Axion-1.0-35B · Hugging Face · 1 src
Comments
via GitHub Discussions