AI21 Labs Jamba2 models: a hybrid built for long documents

AI21 ships the Jamba family, notable for a hybrid architecture that mixes transformer and state-space layers to handle long inputs more efficiently than attention alone. Jamba2 Mini and Jamba2 3B appear in this rating.

Token prices are not published publicly for these entries, so the cells are empty. The interest here is architectural: if long-document throughput is your bottleneck, the efficiency claims are worth testing on your own workload.

• Benchmark against a comparable model from a larger vendor before committing production traffic — line-ups this size get less independent testing.
• Check the context window against your longest input — the largest here is Jamba2 Mini at 256K tokens.
• Request current rates directly: 2 of the listed models publish no public per-token price.
• Confirm support terms and deprecation policy, which smaller vendors document less thoroughly.