Jamba 1.0 — first production hybrid Mamba
AI21 Labs' Jamba (March 2024) was the first production-grade Mamba-based model to ship at scale. 52B total parameters / 12B active via Mixture-of-Experts. The architecture: 7 Mamba layers + 1 attention layer per block, with MoE every 2 layers. Effective 256K context. Apache 2.0 license, open weights.
Jamba was important not because it was the best model in the world (it wasn't) but because it proved hybrid SSM-attention could be productionized. Before Jamba, hybrid architectures were research papers. After Jamba, they were enterprise procurement targets.
Jamba 1.5 — the validation
Jamba 1.5 (August 2024) shipped two versions: Large (94B/398B MoE) and Mini (12B). The Mini is the version most teams will actually evaluate — manageable parameter count, well-documented, available as open weights with commercial license.
The headline benchmark result for Jamba 1.5 was RULER at 256K context. RULER is a long-context benchmark designed to be hard to game — it tests whether the model actually uses the full claimed context window, not just whether the model can technically accept that many tokens. Jamba 1.5 was the only model to achieve effective 256K context on RULER at the time of its release. That's a meaningful claim — many "long context" Transformers fall apart on RULER's recall tests at depth.
The deployment story
Jamba 1.5 is available on AWS Bedrock, Azure AI Studio, and Google Cloud Vertex. Production deployment across all three major hyperscalers is the kind of thing that matters for enterprise customers. The hybrid architecture isn't a research curiosity for AI21 — it's their commercial offering.
Key empirical finding from Jamba's design exploration: 1:7 attention-to-Mamba ratio is the optimal point. Higher attention ratios didn't significantly improve quality but did increase cost. Lower ratios started to degrade recall. The result is approximately a 10× KV-cache reduction compared to a pure Transformer of similar effective size.