Summary
Sam Witteveen looks at Ripple 65B Preview, a new open-weight model from Recursive, a lab founded by the team behind Google’s AlphaChip. The 65-billion-parameter dense model was trained from scratch on 12 trillion tokens with up to a million tokens of context, and the lab says it beats every US open-weight model on Terminal Bench 2.1 with a score of 80.52.
The video puts that number in context against Reflection’s 500B model at 80.1 and Meta Muse Spark at 76.4, and examines the caveat in the “US open-weight” qualifier. It notes the reported $32 million training cost and the unusually detailed write-up. Recursive released instruction-tuned BF16 and NVFP4 versions and a base model under Apache 2.0, and plans to offer it on Together, OpenRouter and others.
The analysis covers design choices: a dense rather than mixture-of-experts architecture, 80 layers, sliding-window attention, gated attention, and an OLMo 3-style model ladder used to test changes at small scale before scaling up. It also covers post-training with in-house synthetic data and 1.4 million RL rollouts. The presenter tests the model for agentic coding on a ZGX Fury workstation using the DeepSeek harness.
📺 Source: Sam Witteveen · Published November 05, 2026
🏷️ Format: News Analysis
