Google rolls out AI model "Gemini Pro", "Gemini
Three model sizes: Nano for phones, Pro for Bard, Ultra targeting GPT-4 in 2024.
Google officially launched its Gemini AI model family on December 6, 2023, starting with Gemini Pro integrated into Bard. The medium-sized model competes directly with OpenAI's GPT-3.5, outperforming it in six out of eight benchmarks including MMLU and GSM8K. Gemini Pro is multimodal—handling text, images, audio, video, and code—but currently lacks image generation (though Google's docs confirm the capability exists). The model is available today in over 170 countries via Bard in English, with broader rollout to Google Cloud on December 13.
Beyond Pro, Google teased two larger variants: Gemini Ultra, designed to beat GPT-4 across 30 of 32 benchmarks and achieve human-expert performance on MMLU, will launch in early 2024 as part of a paid 'Bard Advanced' tier. The smallest variant, Gemini Nano (1.8B and 3.25B parameter versions), is optimized for on-device AI on Android, starting with the Pixel 8 Pro for tasks like voice memo summarization. Google acknowledges Gemini still struggles with hallucinations, and its benchmark claims await independent verification. The full Gemini family will roll out to Search, Ads, and Workspace over the next year.
- Gemini Pro outperforms GPT-3.5 on 6 of 8 benchmarks, available today in Bard across 170+ countries.
- Gemini Ultra claims to beat GPT-4 on 30/32 benchmarks (MMLU, multimodal, etc.) arriving early 2024.
- Nano (1.8B/3.25B params) runs on-device on Pixel 8 Pro; all models use Google's custom TPU chips.
Why It Matters
Google's three-tier Gemini strategy threatens OpenAI's lead, bringing multimodal AI to phones, cloud, and search.