AI21 Labs releases Jamba 1.5

AI21 Labs released Jamba 1.5 in two sizes — Jamba 1.5 Mini and Jamba 1.5 Large — both built on the hybrid Transformer-Mamba mixture-of-experts architecture the company had introduced with the original Jamba five months earlier. Large used 94 billion active parameters routed through 16 experts across 72 layers; Mini used 12 billion. Both offered an effective 256,000-token context window, which AI21 said was the largest available among open-weight models at the time, and both were released publicly on Hugging Face under AI21’s own licence.

The architecture’s pitch was efficiency at long context rather than raw benchmark leadership: interleaving Mamba state-space layers with attention layers gave the models a smaller memory footprint per token than a pure-Transformer design of comparable size, which AI21 said translated into markedly lower latency and throughput cost on long inputs. A quantisation technique the company called ExpertsInt8 let the 94B Large model run on a single machine with eight 80GB GPUs even at full 256K context, without AI21 reporting a quality loss from the compression.

Jamba 1.5 arrived as most open-weight competition remained built on conventional Transformer stacks, making it one of the more prominent attempts to bring an alternative, state-space-hybrid architecture to production scale rather than keeping it a research demonstration.