JOURNAL / 2026.11.05
Ripple-65B releases open weights: what a dense model offers agents
Ricursive introduces a 65-billion-parameter coding model with base and instruction-tuned versions. It expands options for deployment and experimentation; size and a terminal benchmark score alone do not establish savings.
Ricursive Intelligence has released Ripple-65B Preview, a general-purpose variant of the internal model used in its chip design platform. It is a dense Transformer trained from scratch, with subsequent training focused on coding agents. The team explains that it chose this architecture to reduce training and serving complexity. I consider this a relevant contribution because it offers another way to build and adapt agents without relying solely on enormous mixture-of-experts models.
The published license is Apache 2.0. The release includes instruction-tuned weights and a base model. That broadens modification and deployment options, although publishing weights does not provide everything needed to repeat training. For a laboratory, being able to preserve a version and study its changes has value even if it never leads a leaderboard.
It helps to clarify what “smaller” means here. A mixture-of-experts model can store many more parameters while activating only a portion for each token. Ripple is dense: reducing the total weights does not establish that it uses less computation per token than those competitors. The memory needed to store a model, the work of generation, and conversation speed are related but distinct properties. Comparing them through a single size figure can lead to buying the wrong hardware.
The local version changes the question
The NVFP4 variant is marked experimental. It quantizes MLP layers while keeping attention in BF16. Calculating its memory as though every weight occupied four bits would therefore give a misleading estimate. Beyond weights, running an agent requires memory for context and resources for its tools. Fitting a model is the start of an evaluation: we still need to know how long it takes to complete work and what the quantized version loses.
There is also a hosted route: OpenRouter lists it as released on November 4, with tool calling and approximately one million tokens of context. That listing confirms a documented offering without validating its results. An API integration allows teams to explore behavior before preparing their own deployment; they would then need to check that the local version preserves the quality and limits their project needs.
The model card reports 80.5 on Terminal-Bench 2.1, against the published 80.1 for Reflection Beam. These are developer figures: the comparison mixes official results with internal evaluations and warns that changing the evaluation agent changes scores. I would not read those decimal differences as a stable advantage. Results are presented in BF16 at maximum effort; they do not establish NVFP4 performance either.
Terminal-Bench measures task resolution in a terminal environment. Its authors corrected 28 tasks when moving to version 2.1, addressing external dependencies, resources, or mismatches between instructions and tests. That history matters: even executable checks need maintenance. The benchmark provides evidence about one type of work without establishing that an agent can maintain any repository or work without supervision.
Another artifact deserves attention: Ripple-65B Base Preview, the checkpoint before the stages following pretraining. It allows adaptation experiments with a shared starting point. Comparing it with the instruction-tuned version would show the combined effect of several stages; attributing an improvement exclusively to reinforcement learning would require more intermediate checkpoints or controlled experiments.
My assessment is favorable to that research opportunity. I would rather see what useful adaptations a community achieves with these artifacts before turning a narrow terminal advantage into a general proclamation. For teams constrained by memory or access to services, Ripple opens a concrete option to investigate. Adoption will depend on the quality, time, and cost of their complete workflow. The weights make it possible to conduct that comparison themselves.
Primary sources consulted on November 5, 2026. We have not run Ripple or reproduced its evaluations. The implications and experimental proposals are our own analysis.