AI21Labs released Jamba 1.5 Mini & Large models, MoE architecture.
- permissively licensed models with 256K context length.
- multilingual support with JSON output and function call capabilities.
- Jamba 1.5 Large: 94B active parameters (398B total).
- Jamba 1.5 Mini: 12B active parameters (52B total).
- Arena Hard scores: 65.4 (Large), 46.1 (Mini).
- MMLU scores: 81.2 (Large), 69.7 (Mini).
- 2.5x faster inference with similarly sized models.
- enables tool use and feedback integration.
- hybrid transformer/Mamba architecture for strong quality and document understanding.

AI21Labs released Jamba 1.5 Mini & Large models, MoE architecture.
- permissively licensed models with 256K context length.
- multilingual support with JSON output and function call capabilities.
- Jamba 1.5 Large: 94B active parameters (398B total).
- Jamba 1.5 Mini: 12B active parameters (52B total).
- Arena Hard scores: 65.4 (Large), 46.1 (Mini).
- MMLU scores: 81.2 (Large), 69.7 (Mini).
- 2.5x faster inference with similarly sized models.
- enables tool use and feedback integration.
- hybrid transformer/Mamba architecture for strong quality and document understanding.

https://huggingface.co/collections/ai21labs/jamba-15-66c44befa474a917fcf55251

AI21Labs' Jamba 1.5 Mini & Large models offer impressive capabilities with high parameter counts and efficient inference speeds. The hybrid architecture ensures strong quality and document understanding. Great for multilingual support and feedback integration