DeepSeek's MoE model with 552B total parameters and 8-16B active per token. 1M-token context, MIT license: free commercial use. Built for online serving, too large for local use on a single PC.
Ripple-65B Preview
Released on Nov 19, 2026
64.8B-parameter dense model from Ricursive Intelligence, Apache 2.0: pretrained from scratch on 12T tokens, up to 1M context, thin-and-deep architecture (d_model 5120, 80 layers). Aimed at agentic coding: 80.5 on TerminalBench 2.1, ahead of open-weight MoE up to 10x larger. NVFP4 checkpoint (MLP 4-bit, attention BF16) designed for a single 128 GB unified-memory workstation (DGX Spark), ~19.5 tok/s single request.
Features
- 64.8B dense
- Apache 2.0
- 1M context
- TerminalBench 2.1: 80.5
- NVFP4 for 128 GB unified
- Stock SGLang setup on DGX Spark
Version history
- 65B PreviewLatest
Prima release: instruct BF16 e NVFP4, base BF16, DFlash2 draft.
Tags
Useful links
Discover Ripple-65B Preview
Visit the official site to download, learn more and start using Ripple-65B Preview.
Related resources — Models
The most advanced open image model: excellent prompt adherence, solid anatomy, full fine-tuning. The standard for serious work.
Very few steps, high quality, and strong text rendering. The distilled FLUX variant, built for speed.

