TEL AVIV, Israel, Aug. 22, 2024 /PRNewswire/ — AI21, a trailblazer in the development of foundational models and AI systems for enterprises, has unveiled two groundbreaking additions to its Jamba model family: Jamba 1.5 Mini and Jamba 1.5 Large. These models promise unparalleled quality and low latency, furnished with the largest context windows currently available.
With their revolutionary architecture, the new Jamba models eclipse others in their class, outperforming even behemoths like Llama 8B and 70B. Building on the success of their predecessors, these enhanced models signify a substantial advancement in long-context language models, offering unmatched speed, efficiency, and performance across various applications.
AI21’s innovative approach integrates the strengths of Transformer and Mamba architectures, a hybrid strategy that addresses the weaknesses of each. This combination facilitates high-quality, accurate responses while maintaining exceptional efficiency, even with extensive context windows—a feat that traditional Transformer models struggle to achieve.
The zenith of this architectural innovation is Jamba 1.5 Large, a sophisticated Mixture-of-Experts (MoE) model with a staggering 398 billion total parameters and 94 billion active parameters. Designed to excel in complex reasoning tasks, Jamba 1.5 Large is the crown jewel of the Jamba family, embodying unparalleled quality and efficiency.
Alongside this, AI21 introduces Jamba 1.5 Mini, a polished and enhanced spin on the original Jamba-instruct model. It delivers expanded capabilities and superior output quality, and is meticulously crafted with developer-friendliness in mind. Optimized for creating Agentic AI systems, the model supports features like function calling, tool use, JSON mode, structured document objects, and citation mode, making it a versatile solution for developers.
What sets the Jamba models apart is their ability to utilize a true context window of 256K tokens, the largest available under an open license. Unlike many other long-context models, Jamba models fully exploit their declared context window as demonstrated by the new RULER benchmark, which evaluates models on tasks like retrieval, multi-hop tracing, aggregation, and question answering. In these areas, Jamba models shine, delivering consistently superior outputs with an impressively high effective context length.
In rigorous end-to-end latency tests, Jamba 1.5 Large exhibited the lowest latency rate, significantly outperforming competitors like Llama 3.1 70B, Llama 3.1 405B, and Mistral Large 2. Its performance was particularly notable in large context windows, clocking speeds twice as fast as competing models. Similar efficiency advantages were observed when Jamba 1.5 Mini was pitted against Llama 3.1 8B, Mistral Nemo 12B, and Mistral-8x7B.
“We believe the future of AI thrives on models that genuinely leverage extensive context windows, particularly for complex, data-intensive tasks. Jamba 1.5 Mini and 1.5 Large offer the longest context windows in the market, pushing the boundaries of what’s possible with LLM-based applications,” said Or Dagan, VP of Product, Foundation Models at AI21. “Our breakthrough architecture enables Jamba to process vast amounts of information with lightning-fast efficiency, positioning it as the optimal foundation model for developers and enterprises.”
In a significant move, AI21 has partnered with industry giants like Amazon Web Services (AWS), Google Cloud, Microsoft Azure, Snowflake, Databricks, and NVIDIA to ensure seamless deployment and integration of the Jamba models within secure, enterprise-ready environments. Additionally, the Jamba models will be available on popular platforms such as Hugging Face, Langchain, LlamIndex, and Together.AI.
AI21 is also collaborating with Deloitte to bring hyper-customized training solutions and private deployment capabilities to enterprise clients, underscoring the importance of these innovations in meeting diverse business needs. “AI21’s ability to offer private deployments and tailored training solutions is immensely valuable to our clients,” said Jim Rowan, principal and Head of AI, Deloitte Consulting LLP. “Together, we aim to deliver cutting-edge AI capabilities and customized solutions that drive significant value.”
Furthermore, the Jamba models have gained recognition from leading independent benchmarking sites like Artificial Analysis and LMSYS Chatbot Arena. “Jamba 1.5 Mini and Large from AI21 Labs showcase remarkable advantages for inference workloads with long input prompts,” praised Micah Hill-Smith, Co-founder & CEO of Artificial Analysis. “In performance tests with prompts exceeding 10,000 tokens, Jamba 1.5 Mini achieved output speeds significantly higher than the median, affirming both models’ leading performance.”
Through these partnerships and innovations, AI21 is democratizing AI, making it more accessible, scalable, and transformative for a wide range of industries. AI21’s mission to create real-world value by designing enterprise-purposed AI systems continues to drive its leadership in the industry. For detailed information on deploying AI21’s Jamba AI models within your organization, visit https://www.ai21.com/jamba. AI21 also provides custom private deployments, hands-on management, continuous pre-training, and fine-tuning to meet specific enterprise requirements.
About AI21:
Founded in 2017, AI21 specializes in creating foundational models and AI systems that fast-track the adoption of generative AI in production environments. With substantial investments from NVIDIA, Intel, Google, and others, AI21 has been at the forefront of bringing generative AI to the masses. Their AI21 Platform empowers businesses to build their own generative AI-driven applications and services. For further information, visit www.AI21.com.