Model
Inference & Serving
Jamba 1.5
1Hybrid 2transformer-Mamba 3model 4with 5massive 6context
the six Ws · specification
W1
Who
AI21 Labs developed Jamba as a hybrid Transformer-Mamba large language model family.
W2
What
Jamba 1.5 combines transformer and state-space-model layers in a mixture-of-experts architecture with a 256K context window.
W3
Where
Available as open weights on Hugging Face under ai21labs and via AI21's hosted API.
W4
When
Jamba v0.1 released March 2024, with Jamba 1.5 Mini and Large following in August 2024.
W5
Why
Built to deliver long-context efficiency and throughput beyond pure transformer architectures.
W6
With
Runs via Hugging Face transformers or vLLM, or through AI21's Studio API.
W7
Watch
Open weights under the permissive Jamba Open Model License allowing commercial use; hosted API requires an AI21 key.
alternatives
works with
hybrid-architecturestate-space-modellong-contextmixture-of-expertstransformer-SSM