AI21 Models

AI21 develops the Jamba family of language models. Its Jamba2 releases use a hybrid Transformer–Mamba architecture and focus on document grounding and instruction-following. The checkpoint you choose determines the runtime and resource requirements; the family name alone does not.

Three starting points

Jamba2 Mini has 52B total parameters and 12B active parameters. It is a candidate for a document assistant that needs to follow a supplied policy or technical reference. Jamba2 3B offers a smaller checkpoint for tighter deployment budgets. Jamba Reasoning 3B has a different reasoning-oriented interface and serving configuration.

Grounding is a testable requirement

Use real documents with version conflicts, similar names and questions whose answers are absent. Require a passage reference and inspect it. A useful assistant should distinguish the current policy from an obsolete paragraph and leave an answer unresolved when the evidence is missing.

Record format compliance as well as answer quality. An extracted record that fails your schema cannot enter the next stage of a workflow, even if its explanation sounds sensible.

Weights and hosted services

These three publisher checkpoints use Apache-2.0 licensing. That licence describes the distributed weights; it does not include free hosted inference. Review the current model card for runtime support and confirm any managed endpoint independently. An old Studio example is not evidence that the same model identifier remains available.

Compare the two Jamba2 sizes using the same documents, output limits and acceptance criteria. Measure memory, accepted-answer cost and the corrections required from a person.

Publisher resources

Current publisher documentation