Reasoning weights with a permissive license.

The human problem

Teams need models they can inspect, adapt, and deploy under clear terms without sending every task to a hosted service.

The prior constraint

Many capable models expose only an API or weights with restrictions that complicate reuse.

AI’s actual role

The models generate text, reason over prompts, and call tools in developer-built systems.

The documented result

Weights and model cards are publicly available in three sizes under Apache-2.0. Capability and benchmark claims are IBM-reported, not deployment outcomes.

Why it may matter

A permissive release gives researchers and organizations more practical control over deployment and adaptation.

Limitations

Released weights do not disclose every training datum or guarantee full reproducibility.

The 30B model still requires substantial hardware.

Benchmarks do not establish reliability in a particular real-world workflow.

Unresolved questions

How do the sizes compare under independent evaluation?

What training-data documentation and opt-out mechanisms are available?

Which tool-use failures appear in deployment?