Mixtral 8x22B

• Released: 2024
• Physical: 141B parameters
• Active: ~39B per token
• Performance: Excellent on benchmarks
• Context: 32K tokens extended
• Use: High-performance inference