Jamba

AI21 Labs builds Jamba by interleaving Mamba state-space layers with a few attention layers, plus mixture-of-experts blocks in larger models, so long prompts do not balloon the attention cache as they do in a pure Transformer. Every release lists a 256K-token window. The original Jamba v0.1 base model (12B active, 52B total parameters) and the January 2026 Jamba2 releases, a 3B model and a Mini of the same 12B/52B shape, are Apache-2.0. The 1.5, 1.6 and 1.7 Mini and Large instruct models, including the 1.5 Mini linked here, sit behind a click-through gate under AI21's Jamba Open Model License, which bars commercial use by organizations above 50 million US dollars in annual revenue. Jamba Reasoning 3B, also in GGUF form, is small enough for laptops and phones. Mini-class models need data-center GPUs and run through vLLM (0.12+ for Jamba2 Mini, with expert int8 quantization) or Transformers plus the causal-conv1d and mamba-ssm kernels; per the v0.1 card, 8-bit weights let one 80 GB GPU hold about 140K tokens of context. AI21 also sells hosted access. Best suited to teams doing grounded question answering or RAG over long documents.