Models & Architectures

Hybrid Model Architecture (Transformer-Mamba)

AI models that blend standard Transformer layers with a state-space design like Mamba to handle long context more cheaply.

Also known as: state-space model,Mamba,transformer-mamba hybrid

Almost every well-known AI model — GPT, Claude, Gemini, Llama — is built purely on the Transformer architecture. A hybrid model, like AI21's Jamba, blends Transformer layers with Mamba, a "state-space model" that carries forward a compact running summary instead of re-examining every previous token. Transformers get more expensive the longer the input gets; Mamba layers stay cheap and flat at long context, at some cost to precise recall. Blending both lets a hybrid model handle very long documents — entire contracts or codebases — with a lower memory footprint and lower serving cost than a comparable pure-Transformer model, without fully sacrificing precision.

Read the full guide

Tools that use this