AI21 LabsJamba2

Jamba2 Mini

Jamba2 Mini is the larger of the two January 2026 Jamba2 releases. Its mixture-of-experts design activates 12 billion of 52 billion parameters for each token.

How Jamba2 Mini was trained

The model starts from the Jamba 1.5 pretraining stage. AI21 added 500 billion tokens of mathematics, code, web text and long documents. Later stages included supervised tuning, preference optimization and reinforcement learning. State passing during training helped the model handle long sequences.AI21-Jamba2-Mini official model card

Which API name selects this model

AI21 maps jamba-mini and jamba-mini-2 to jamba-mini-2-2026-01. These are aliases for this release, not additional models.AI21 Jamba versions, aliases and API deprecations

Practical guidance. Pin the dated name when comparing outputs across deployments. A moving alias can select a different release after the provider updates its service.

Active parameters do not set the memory budget

The public weights use Apache 2.0. AI21 recommends vLLM 0.12 or later for the published serving example and also distributes an FP8 export.AI21-Jamba2-Mini official model card

Practical guidance. The 12-billion active count describes token computation. A self-hosted server still needs to store the model's experts. Measure memory and latency with the intended document length, batch size and weight format.

Specifications & access

Jamba2 Mini specifications
DetailValue and conditions
Launch date
AccessAPI, Downloadable
Model identifiersAI21-Jamba2-Miniai21labs/AI21-Jamba2-MiniAI21-Jamba2-Mini-FP8jamba-minijamba-mini-2jamba-mini-2-2026-01
Context window262,144 tokens

Maximum position count in the published weight configuration; deployed limits depend on the serving setup.

AI21-Jamba2-Mini configuration
Parameters52 billion total; 12 billion activeAI21-Jamba2-Mini official model card
Weight licenseApache 2.0AI21-Jamba2-Mini official model card

Related models

  • Compact dense language model for private document processing and instruction-based tasks.

    Launch date Downloadable
  • Large hybrid mixture-of-experts model focused on answers grounded in supplied documents.

    Launch date API · Downloadable
  • Earlier Mini release with improved source-grounded answers and instruction following.

    Launch date Deprecated API; check access detailsDownloadable