Ripple-65B by Ricursive: how it works and why it could change AI and SaaS

A model with 65 billion parameters, a team of five researchers, and 32 million dollars of development spend. Ricursive Intelligence has just introduced Ripple-65B, an open model that questions how much it really costs to build competitive artificial intelligence.

For the last few years, the AI race has been dominated by one idea: to get smarter models we need more parameters, more infrastructure, and more investment.

On 4 November 2026, Ricursive Intelligence introduced Ripple-65B Preview, a language model that scores 80.5% on Terminal-Bench 2.1, a test of how well agents solve complex tasks with development tools.

The result is competitive with considerably larger models. What is actually interesting is the cost: Ricursive says it built the model for 32 million dollars, about a tenth of the budget it estimates for its main US reference.

Are we looking at a new way to build more efficient AI models? What can change for SaaS companies?

What Ripple-65B is

Ripple-65B Preview is a general-purpose language model from Ricursive Intelligence, tuned especially for autonomous programming, reasoning, and getting work done with tools.

These are its main traits:

It is an open-weight model: the weights are public and can be downloaded from Hugging Face, run on your own infrastructure, or used through inference providers.

The Apache 2.0 license allows commercial applications and custom builds.

Who is behind Ripple

Ricursive Intelligence was founded by Anna Goldie and Azalia Mirhoseini, researchers with experience at Google DeepMind and in building AI systems.

Their aim is more efficient models, by improving the AI and the computing infrastructure together.

Ripple was built by a small research team, but Ricursive is a well-funded company: in January 2026 it announced a 300 million dollar round at a 4 billion valuation.

Where to try Ripple-65B

One advantage of the launch is that the weights are public. API access, though, is still settling.

There are three main paths right now:

1. OpenRouter: Ripple-65B Preview

This is the most direct way to try the model without standing up your own infrastructure. The model card is on OpenRouter.

You can open the card, sign in, and select it in the OpenRouter chat. You can also use its OpenAI-compatible API with the id ricursiveai/ripple-65b-preview.

Important for users in Europe: OpenRouter says Ripple is not yet available through EU in-region routing. That does not mean it is blocked for people in Spain or other European countries. It means processing inside Europe is not guaranteed. For first tests, use non-sensitive data and check that the provider is actually available.

2. Together AI: Ripple-65B Preview

Together AI publishes model details, prices, and API examples.

At the time of writing, the page also shows a Launching soon notice. Check whether the endpoint is live before you try to call it.

3. Hugging Face: Ripple-65B Preview

If you would rather run it locally or on private infrastructure, Ricursive has published the model on Hugging Face in BF16, plus an NVFP4 version that needs less memory.

That path needs the right hardware and people who know how to deploy models.

How Ripple-65B works, and what is different

One of Ricursive's more interesting choices is a dense architecture, instead of Mixture-of-Experts (MoE), which many large open models use.

In an MoE model, only part of the parameters take part in each token. In Ripple, all 65 billion parameters are active.

To offset that cost, Ricursive tuned the architecture with 80 relatively narrow layers, more efficient attention, and training improvements.

The main novelty is not one revolutionary technique. It is the combination of several optimizations.

According to the official technical notes, those changes reached up to 16 times the compute efficiency of their Olmo 3 reference architecture, in code-training experiments at a 4B-parameter scale.

Ripple also reached its Terminal-Bench result with about 1.4 million reinforcement-learning rollouts, roughly 70 times fewer than Reflection Beam, according to Ricursive. A rollout is a practice run: the model attempts a task and is scored on the result.

That does not mean Ripple is 70 times cheaper or faster. Rollouts can cost different amounts.

The more interesting conclusion is different: Ricursive seems to have found ways to improve how the model learns without relying only on more compute.

What Ripple-65B is for

Ripple can be used as a general language model, but it stands out on technical work that takes several steps.

1. Programming and software maintenance

Ripple is tuned for agents that work on code repositories.

An agent could take an incident, inspect the code, find the bug, change the files, and run tests to see whether it fixed it.

Its SWE-bench and Terminal-Bench results show potential for that kind of work.

2. Doing tasks through a terminal

This is probably its most interesting use.

Instead of only generating code, an agent can run commands, install dependencies, edit files, launch tests, and fix errors.

Some Terminal-Bench tasks, for example, ask the agent to train machine-learning models under accuracy and size limits, configure environments, or resolve programming incidents.

To finish the job correctly, the agent has to use a terminal and check the results, not just explain how it would solve the problem.

That connects with something we already covered: MCP versus CLI.

A CLI (the command line) runs commands in a controlled environment. MCP is a standard interface for discovering and calling tools. Both can sit in the same agent architecture, depending on what the product needs.

Ripple-65B against GPT, Claude, and other models

The number that drew the most attention is 80.5% on Terminal-Bench 2.1.

The test scores tasks run through a terminal, including programming, debugging, configuration, and technical operations.

To put it in context, we can set Ricursive's published results next to Terminal-Bench on Vals AI.

The results use Terminus 2 as the reference, but they come from evaluations by different organizations and the setup can differ. The last two competitors are evaluations run by Ricursive.

Ripple does not beat the most advanced commercial models, but it gets close while using far fewer total parameters than other open models.

It is not equally competitive on every test.

Source: Ricursive's model card, which mixes official results and their own evaluations.

The MCP Atlas gap matters especially for business applications.

Ripple scores 71.7%, against 89.2% for Muse Spark Open. Doing well at terminal work does not automatically mean being the best model for connected tools.

Ripple is competitive in a specific category. It is not yet the best AI model for every task.

What Ripple-65B costs: training and inference

This is probably the most disruptive part.

Training cost: 32 million dollars

Ricursive says it developed Ripple for 32 million dollars, about 28.4 million euros.

The figure includes research, experiments, data preparation, evaluations, training, and post-training.

As a historical reference, OpenAI said GPT-4 training cost more than 100 million dollars. That is not a direct comparison: the projects are from different generations, and the figures do not necessarily include the same line items.

Even so, Ripple is an interesting signal that there may be more efficient paths to competitive models.

Inference price, next to other models

On the prices published by OpenRouter, Ripple costs 0.20 dollars per million input tokens and 0.60 dollars per million output tokens.

We can set that next to other models using public rates from Together AI.

Prices published on 5 November 2026. Cache discounts are not included.

Ripple is competitively priced. It is not the cheapest.

Paying less per token is not the same as spending less to finish a task.

What each correctly solved task costs

We can run a simulation with the prices above and the Terminal-Bench scores.

Assume each attempt uses 200,000 input tokens and 50,000 output tokens, for every model.

A simulation with the same token use per attempt and an approximate rate of 0.8888 EUR/USD. These are not costs measured on a real Terminal-Bench run.

Under those conditions, Ripple's cost per success looks very attractive.

Different models can still need different amounts of tokens, steps, and retries to finish the same task.

To know the real efficiency we would have to run the models under the same conditions and measure what each one consumes.

That is what we cover in the token price is not the agent's cost: in business applications, the metric that matters is what it costs to finish the work correctly, not only what it costs to generate tokens.

What Ripple could change in the AI industry

Ripple has just launched, and independent evaluations are still missing, but it points at three trends.

Lower barriers to building competitive models. If the training methods can be reproduced, smaller teams could reach advanced capability without investments on the scale of the big labs.

More specialization. Instead of relying only on ever-larger general models, we could see models tuned for programming, tools, documents, or specific business processes.

More competition in inference. Open models let you use different providers and try private deployments. That can reduce dependence on a single vendor and widen the infrastructure options.

A dense 65B model still needs considerable resources. And the efficiency numbers Ricursive published still need more outside validation.

The interest in Ripple is not that it has proved better than every existing model. It is how close it can get to certain leading capabilities with a relatively small development budget.

What Ripple-65B means for SaaS companies

For software companies, the impact can go beyond the launch of another LLM.

The most obvious consequence is economic: if capable models get cheaper to use, more processes become candidates for automation.

There is also a strategic consequence.

As specialized models appear, picking one provider for every AI need may stop being the most efficient choice.

A SaaS could use a small model to classify requests, another to run technical operations, and a more advanced model for hard problems.

The hard part will be choosing the right model for each task, and managing the tools, the data, the permissions, and the evaluations that work needs.

That is what we cover in the agent harness: a good model is only one of the pieces required to build reliable agents.

This shift can also touch what makes a SaaS defensible.

In our piece on REA, ArtCraft, and reverse engineering with AI we explain how AI is lowering the barrier to building and recreating software features.

Ripple points the same way from another angle: the intelligence needed to automate complex processes could also become more accessible.

In that setting, having a powerful model will not be enough to stand apart.

The advantages will sit more and more in industry knowledge, data, integrations, and the ability to run business processes reliably.

The new frontier may be efficiency

Ripple-65B is not the most advanced AI model in the world. GPT, Claude, and other models still beat it on various tests.

Ricursive has done something worth attention: an open model that is relatively compact, competitive at autonomous programming, with a stated R&D cost that is surprisingly contained.

If the methods behind those results spread, they could speed up new specialized models and lower the barrier to advanced AI.

For SaaS companies, that is an opportunity.

Models will keep getting easier to reach. The real difference will be how we put them into products to solve real problems, reliably and at a cost that works.

At Devic that is the layer we work on: connecting models, tools, and data so AI capability becomes agents that can operate inside SaaS applications.

The future of the software will not depend only on which model it uses, but on what work it can actually finish.