Deca 2.5: Intelligence on Demand

---

Today we are releasing Deca 2.5, our latest generation of models. The series includes three variants—Mini, Pro, and Ultra—each designed for different use cases but sharing a common architectural foundation: Dynamoe, our proprietary approach to sparse, routed inference.

Deca 2.5 addresses what we believe is the core question in AI: the balance between capability and cost. Intelligence that no one can afford is not useful. The relationship between capability and cost determines what developers can actually build in the real world.

The Models

Why This Matters

For the past few years, the AI industry has operated on a simple assumption: the only way to build better models is to use more computing power. Bigger data centers. More GPUs. Massive power bills. The idea is that the company with the biggest supercomputer wins.

This isn't entirely wrong. Scale does matter. Training a model on more data makes it smarter.

But treating raw compute like it's the only tool available ignores a much more interesting question: how much of that compute is actually doing useful work?

In standard AI models (called dense transformers), every part of the model activates for every question. A simple question still engages the entire model's capacity—all the parameters, all the computational overhead. This is inefficient. It is like running a full data center to process a single request.

Some newer models use an approach called Mixture-of-Experts (MoE) to fix this, and they do help. But they are usually stuck inside one giant, rigid model.

Dynamoe

Deca 2.5 is built on Dynamoe. Instead of being one massive, rigid brain, Dynamoe is a network of highly specialized experts and adapters. When you ask a question, Dynamoe instantly figures out exactly which specific experts are needed, connects them, and leaves the rest of the network completely turned off.

If you ask a simple fact, it only activates a few relevant experts. If you ask a complex coding question that requires mixing Python and database knowledge, it activates a different, larger set of experts.

This has several practical consequences:

• Lower inference cost. When you are not activating the full model for every token, you are not paying for the full model for every token. This is the direct mechanism by which Deca 2.5 achieves its price point.
• Lower latency under load. Sparse activation means less computation per forward pass, which means faster responses, particularly under high concurrency.
• Better specialization. Individual experts and adapters can develop deeper competence in their domains than a single dense model can, because they are not forced to share capacity with every other domain.
• Compositional flexibility. New experts and adapters can be added to the system without retraining the entire model. This allows us to extend the model's capabilities incrementally—adding domain expertise, supporting new languages, adapting to new tasks—without the enormous cost of a full training run.

Capability

We want to be straightforward about how Deca 2.5 compares to the rest of the industry.

Deca 2.5 Pro is highly competitive with top-tier models on the tests that actually matter: complex reasoning, writing code, understanding long documents, and following instructions. It doesn't win every single benchmark. No model does. But for most real-world work, it performs at a level that is indistinguishable from models that cost five times as much.

Deca 2.5 Ultra (in beta) goes even further. Its massive 1M token context window lets it handle tasks that other models simply can't do: reading an entire codebase, summarizing hundreds of legal documents, or keeping a long-running agent on track for hours.

Deca 2.5 Mini handles everyday tasks: sorting data, summarizing emails, simple classification. It is not the most capable model, but it is fast and inexpensive. For the majority of AI use cases, this is sufficient.

We publish test scores because they are a useful baseline, but we encourage developers to test Deca 2.5 on their own actual work. The only benchmark that truly matters is whether the model gets your specific job done.

Pricing Philosophy

Our pricing is a direct result of our architecture. Because Dynamoe uses far less computing power per word, we can charge significantly less than our competitors.

We aren't artificially lowering our prices just to grab market share. Our prices are based on our actual costs, meaning they are sustainable. They won't suddenly jump up next year.

We believe pricing dictates what developers are able to build. When an AI model costs $30 per million tokens, developers are careful. They limit how often they use it. But when it costs $2.00, they build differently. They let AI agents run longer. They add more review steps. They build entirely new types of software that simply wouldn't have been profitable before.

Cheaper intelligence isn't just about saving money. It's about unlocking new ideas.

Availability

Deca 2.5 Mini and Pro are available today through the Deca API, the Playground, and Deca Chat.

Deca 2.5 Ultra is available in private beta. You can request access through the Playground.

All new accounts receive $2.50 in free credits—enough for meaningful experimentation across any of our models.

What Comes Next

Deca 2.5 is not an endpoint. It is the current expression of an approach we believe in deeply: that the future of AI belongs to architectures that achieve more with less.

We are already working on what comes next. We will share more when we are ready—not when we need a press cycle, but when we have something worth showing.

In the meantime, we invite you to build with Deca 2.5. Try it on your hardest problems. Push it. Tell us where it falls short. We are listening.

---