Google Updates Gemini 1.5, Reveals New Models
Google has not released Gemini 2, but instead updated the technical report for Gemini 1.5 in May 2024. This update includes new benchmarks, technical information, and details on Gemini Flash and the upcoming Gemini 8B. Gemini 1.5 Pro (May Update): 🧮 Sparse multimodal, multilingual Mixture of Experts (MoE) 🎥 Supports text, audio, image, and video inputs 🥇 78.1% win rate vs Gemini 1.5 (Feb) & 77.8% vs Gemini 1.0 Ultra 🌟 Outperforms 1.0 Ultra on 16 of 19 text benchmarks and 18 of 21 vision benchmarks 🧑🏻💻 Improved performance in math, coding, and multimodal benchmarks 🖼️ Enhanced text and vision task understanding 🧠 Achieves 85.9% on MMLU (5-shot) and 91.7% with majority voting 🔍 Supports 1 million token context, 2 million via waitlist 🌍 Strong in low-resource language translation 💰 Costs $3.50 per 1M tokens (prompts up to 128K tokens), $7.00 for above; $10.50 per 1M tokens (prompts up to 128K tokens), $21.00 for above Gemini 1.5 Flash (New): 🥈 78.9% on MMLU & 82.0% win rate vs Gemini 1.0 & 46.7% vs Gemini 1.0 Ultra 🧪 Dense Transformer decoder model, trained with online distillation from 1.5 Pro 🏎️ 3x faster than Gemini 1.5 Pro, ~4x faster than GPT-4 Turbo, ~1.5x faster than Claude Haiku 🎥 Supports text, audio, image, and video inputs 🔍 Supports 1 million token context, 2 million via waitlist 🤯 10x cheaper than Gemini 1.5 Pro 💰 Costs $0.35 per 1M tokens (prompts up to 128K tokens), $0.70 for above; $1.05 per 1M tokens (prompts up to 128K tokens), $2.10 for above Gemini 1.5 8B Flash (Upcoming): 🧩 Currently in training 🚀 Handles contexts over 1M tokens with maintained performance 🌍 Scores 68.1% on MMLU, matching Meta Llama 3 8B These advancements promise to inspire new research in AI, though it's noted that big tech companies aren't sharing all their experimental details.