← All models

Jamba2 Mini

AI21 Labs · 2026-01-08 · 51.6B parameters

Jamba2 Mini is AI21 Labs' open-weight hybrid Mamba-Transformer mixture-of-experts instruction model, announced on January 8, 2026; its checkpoint contains 51,570,323,328 parameters, which AI21 describes as 52B total and 12B active, and it supports a 262,144-token (256K) context. AI21 built it from Jamba 1.5 pre-training with 500B tokens of mid-training, a Mamba state-passing phase, cold-start SFT, DPO, and on-policy reinforcement learning, targeting grounded question answering and instruction following without the thinking-token overhead of reasoning models. Its 32-layer architecture combines 28 Mamba layers and four GQA attention layers with 16 experts and top-2 routing; the weights are released under Apache 2.0, and AI21's API listed the snapshot as jamba-mini-2-2026-01.

Benchmark scores