AI21 Labs has launched Jamba2, a series of open-source language models designed for enterprise use, including the Jamba2 Mini with 52 billion parameters. This model is optimized for precise question answering and offers a memory-efficient solution with a 256K context window, making it suitable for processing large documents like technical manuals and research papers. Jamba2 Mini excels in benchmarks such as IFBench and FACTS, demonstrating superior reliability and performance in real-world enterprise tasks. Released under the Apache 2.0 License, it is fully open-source for commercial use, offering a scalable and production-optimized solution with a lean memory footprint. Why this matters: Jamba2 provides businesses with a powerful and efficient tool for handling complex language tasks, enhancing productivity and accuracy in enterprise environments.
Sopro is a compact text-to-speech model with 169 million parameters, designed for real-time applications and capable of zero-shot voice cloning. It supports streaming and can generate 30 seconds of audio in just 7.5 seconds on a CPU, requiring only 3-12 seconds of reference audio for effective voice cloning. While it is not state-of-the-art and occasionally struggles with voice likeness, Sopro is a notable achievement given its development on a single L40S GPU and limited resources. The model is available under the Apache 2.0 license, although it currently supports only English due to data constraints.