Jamba
First production-grade hybrid Mamba-Transformer-MoE with 256K context
About
Jamba, released March 28, 2024, was the first production-scale model built on Mamba: a hybrid that mixes Mamba structured state-space layers with Transformer attention and mixture-of-experts routing, totaling 52 billion parameters with 12 billion active. The payoff is long-context efficiency — a 256K token window, with up to 140K fitting on a single 80GB GPU, and team-reported 3x throughput on long contexts versus Mixtral 8x7B. Weights are Apache 2.0. The catch is operational: optimized inference needs the mamba-ssm and causal-conv1d kernels, and the hybrid architecture saw less community tooling than pure Transformers. AI21 later extended the line with Jamba 1.5 Mini and Large and a small Jamba Reasoning model.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: matches or beats models in its size class on standard benchmarks; 3x throughput on long contexts vs Mixtral 8x7B (team-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: Released 2024-03-28 per AI21 and press; Apache 2.0 confirmed; 52B total / 12B active, hybrid Mamba-Transformer-MoE, 256K context (team specs); Throughput claims (3x vs Mixtral 8x7B) are team-reported; Operational complexity (mamba-ssm kernels) noted as limitation.