Lightweight, fast and cost-efficient while featuring multimodal reasoning and a breakthrough long context window of up to one million tokens.