Latest in AI
Ripple-65B Preview: API Pricing, Benchmarks and Its Dense Architecture
Ricursive's Apache-licensed coding model combines 65 billion active parameters with a million-token context window. Its benchmark results need a closer look.
Ricursive Intelligence released Ripple-65B Preview on November 4, a dense open-weight model aimed at reasoning and agentic coding. It supports up to one million tokens of context and is available through hosted APIs as well as downloadable weights. The interesting question is how its relatively compact, fully active architecture performs on the work a coding agent repeatedly does.
Ripple-65B weights and API pricing
The model card lists an Apache 2.0 license, BF16 and NVFP4 instruction-tuned weights, and a BF16 base model. It has 65 billion parameters and was pretrained on 12 trillion tokens.
At publication, OpenRouter's listing shows a GMICloud route with these terms:
| Item | Listed value |
|---|---|
| Input tokens | $0.20 per million |
| Output tokens | $0.60 per million |
| Cache reads | $0.05 per million |
| Listed model context | 1M tokens |
| EU in-region routing | Not currently available |
OpenRouter currently lists a one-million-token context window. Check the provider route and its request limits before building around that maximum, since endpoint settings can change. A long input limit also does not establish how well a model finds and uses evidence across that input.
What the Ripple benchmarks show
Ricursive’s model card reports 80.52 on Terminal-Bench 2.1. Its comparison mixes published scores with in-house evaluations and acknowledges that agent harnesses can shift results. These are vendor-reported findings, not an independent assessment.
The same table gives Ripple 76.5 on SWE-bench Verified, below Reflection Beam's 81.6. A strong terminal-agent result therefore should not become a claim that Ripple leads every coding task.
For a developer, the useful test is closer to the work an agent will repeatedly perform: find a bug in your repository, make a change, run the tests and recover when something fails. Record the full attempt, including retries. A low token price is attractive, but repeated unsuccessful runs can erase that advantage.
Why the dense architecture matters
The launch explanation describes a dense architecture, with all parameters active, rather than a mixture-of-experts design. Smaller total parameter counts can reduce weight-storage requirements, but do not by themselves prove cheaper inference than a larger model that activates only a subset of its weights.
This is the distinction to keep in mind when reading size comparisons. Deployment cost also depends on precision, context, batch size and the hardware serving the model. A local checkpoint gives teams control over deployment, but it does not remove those engineering choices.
For teams considering self-hosting, start with the smallest realistic workload and measure memory use as the conversation grows. Loading weights successfully is only the first step; the agent still needs room for its context and useful response speed.
Who should try Ripple-65B Preview
Ripple looks worth evaluating for developers who want an Apache-licensed coding model and can compare it against an existing agent on their own tasks. The preview label makes a staged trial more sensible than an immediate wholesale migration.
The next useful result will be reproducible performance at a known serving cost. That is what turns an impressive model release into a dependable option for a product team.
I’m Paras. I work at the intersection of AI, engineering and distribution.
I help AI founders explain technical launches clearly, so developers understand what changed and when it matters.