65B Dense Model Ripple-65B Preview Released Under Apache 2.0
Posted As of
Ricursive Intelligence has released the 65B dense model Ripple-65B Preview under Apache 2.0. It supports a context of up to about 1 million tokens, and all benchmarks are the developer's own figures.
What happened
Based on the model card posted on Hugging Face.
- Model: Ricursive Intelligence's Ripple-65B Preview. It is a dense transformer in which all 65B parameters are active, not a mixture-of-experts (MoE) structure. 80 layers, hidden size 5120. (Official)
- License: Apache 2.0. (Official)
- Context: Up to 1,048,576 tokens. The card states that pretraining used an 8K length. (Official)
- Training: Trained from scratch on about 12 trillion tokens, mainly English text and code. (Official)
- Distribution: This repository is the BF16 instruction-tuned version. Separately there are an NVFP4 quantized version, a BF16 pretrained base version, and a DFlash2 variant. (Official)
- Hardware mentions: The card says the model fits on consumer hardware such as the NVIDIA DGX Spark, but does not specify the precision basis. There is also a figure of about 250 tokens/s on 2×B200 at batch size 8, which is the pre-quantization value. (Official, developer-reported figure)
- Benchmarks (developer-reported): Terminal-Bench 2.1 80.5, SWE-bench Verified 76.5. (Official, self-claimed)
We could not confirm the release date. The model card has no date, and the citation information only says 2026. Reports differ, giving November 4 to 7.
Not in the card: whether llama.cpp, vLLM, or MLX support exists, required memory and measured speed on Mac, independent evaluations, and a paper.
Our take
Because it is a 65B dense model, its memory demands are higher than those of an MoE with fewer active parameters. By simple calculation, the 4-bit weights alone are about 33GB, and context and runtime overhead are added on top (calculation, not measured). Mac Studio-class unified memory or DGX Spark are candidates, but we have not verified this on our channel.
The model repository carries a custom-code tag, so we cannot tell whether it runs directly on llama.cpp or MLX right now. If it is confirmed to run, it is worth considering as a measurement candidate under the same conditions.
Sources and related posts
- Ripple-65B Preview 모델 카드 (Hugging Face) Primary source