Jamba

Jamba is a family of open foundation models developed by AI21 Labs, an Israeli company founded in 2017, designed specifically for enterprise efficiency. The models target organizations that require high accuracy, low latency, and data privacy across regulated industries such as finance, healthcare, defense, and technology. Jamba excels at tasks demanding reliable outputs and secure deployment, offering both cloud and on-premises options to keep proprietary data under the organization's control. The family includes Jamba 1.5 Mini (52B parameters, 12B active), Jamba 1.5 Large (398B parameters, 94B active), Jamba2 3B, Jamba2 Mini, and Jamba Reasoning 3B, each tailored for different workload scales from on-device applications to enterprise-wide reasoning.

Jamba's key differentiator is its hybrid Mamba-Transformer architecture, which combines the efficiency of state-space models with the quality of transformers to deliver the fastest processing speeds on the market while maintaining top accuracy. All models support a 256K token context window, enabling them to handle lengthy documents like financial records, contracts, or entire knowledge bases in a single pass. The architecture also yields significant cost efficiency: Jamba achieves higher throughput per compute unit compared to pure transformer models, especially over long contexts. The models officially support nine languages: English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic, and Hebrew. Additional capabilities include structured JSON output, function calling, and digesting document objects, making them suitable for complex enterprise workflows.

In benchmark comparisons, Jamba 1.5 Mini scored 46.1 on the Arena Hard benchmark, making it the strongest open model in its size class, outperforming comparable models from Mistral, Meta, and Cohere. Jamba 1.5 Large outpaces both Llama 3.1 70B and 405B on key metrics. Jamba 1.6 further raises the bar with industry-leading results across enterprise use cases, achieving cited responses with over 90% consistency across long contexts and improving data classification by 26 percentage points over previous models. The models are available on cloud partners including Google Cloud Vertex AI, Microsoft Azure, and NVIDIA NIM, with upcoming support for Amazon Bedrock, Databricks Marketplace, Snowflake Cortex, and Together.AI, plus private on-prem and VPC deployment.

Despite its strengths, Jamba has notable trade-offs. The hybrid Mamba-Transformer architecture is more complex than standard transformers, which may increase the learning curve for teams unfamiliar with state-space models. Public documentation is less extensive than that of competitors like Meta's Llama or Mistral, potentially slowing integration. The pricing for the AI21 platform is $0.20 per 1M input tokens and $0.40 per 1M output tokens for both Mini and Large models, which is competitive but may still be higher than some open-weight alternatives when self-hosted. Additionally, while the models are open with weights available for download, the full ecosystem of community tools and fine-tuning recipes is not as mature as that for Llama or Mistral, requiring more in-house expertise for customization.

$0.20 per 1M input tokens and $0.40 per 1M output tokens for both Jamba 1.5 Mini and Jamba 1.5 Large on the AI21 platform; cloud provider pricing varies.