Mantus 2 improves long-context reasoning

Mantus 2 ships with a better story for people who give it long, messy, real-world documents. The improvement is not a bigger context window — it is a better grasp of what matters inside the window.

This release is the first where our reasoning research made it directly into the product, and I want to show what that looks like.

The problem with long context

A context window of hundreds of thousands of tokens sounds like the ability to “read everything.” The uncomfortable truth is that reading everything and understanding everything are different things. When you paste in a fifty-page document and ask a question that requires stitching together two facts from page three and page forty, the model has to carry a conclusion across a lot of unrelated material.

Our research on compositional reasoning under distraction showed this is exactly where models are fragile: accuracy on multi-step reasoning falls by nearly a third when unrelated content sits between the steps. For anyone working with long documents, that fragility is the whole experience.

What changed

Two fixes, from the paper and from the engineering team:

Retrieval-then-reason. When a question spans a long document, Mantus 2 pulls the relevant passages together before it reasons, rather than reasoning over the whole stream. The model reads like a careful person: locate the relevant parts, then think. This is the product pattern our research recommended, and it recovers most of the accuracy lost to distraction.

Explicit intermediate restatement. Mantus 2 is trained to restate what it has concluded before moving to the next step. It is a small behavioral change with an outsized effect: carrying a conclusion forward explicitly keeps it from being diluted by everything in between.

The result

On our internal long-document evaluation, Mantus 2 answers cross-document questions correctly about a fifth more often than Mantus 1. Users with long-document workflows are the ones who will feel it most.

The deeper point is the loop. A paper from the reasoning team — written to understand a real failure mode — became a product change within one release cycle. That is how we want research and product to work at Curos: questions from real use, answered with real methods, shipped back to the same users.

All news →