Google Bard Gets Gemini Pro in 40 Languages and Hits #2 on Popular Benchmark Leaderboard
And, it now has image generation
Google Bard received a global update today that extends the Gemini Pro large language model (LLM) feature to 40 languages. The company says this is now available in 230 countries. It also added image generation to Bard and highlighted a new quality recognition from a leading benchmark.
The Model
Gemini Pro is currently Google’s most advanced LLM. The more advanced, multimodal Gemini Ultra model is expected to be launched in “early” 2024. The Pro model was rolled out for English for U.S. users in December. Today’s announcement extends that to all other markets supported by Bard. Users can expect more accurate and higher-quality responses than the earlier version of Bard, which employed the PaLM LLM.
Last December, we brought Gemini Pro into Bard in English, giving Bard more advanced understanding, reasoning, summarizing and coding abilities.
Adding Visuals
The upgrade also brings the Imagen 2 text-to-image generation feature to Bard. The image at the top of this article was created using Bard, and you will find it similar to other text-to-image generators. It generates two images per prompt, though a single button click automatically generates two more.
For an extra creative boost, you can now generate images in Bard in English in most countries around the world, at no cost. This new capability is powered by our updated Imagen 2 model, which is designed to balance quality and speed, delivering high-quality, photorealistic outputs.
The most notable element of this feature is that AI image generation is free and fairly fast. Bing Create is also free, but the most popular solutions, such as Midjourney and DALL-E through the OpenAI playground, carry a cost after an initial free trial. Other services offer image generation as part of a monthly subscription plan. Upscaling isn’t available yet, though when I asked to upscale an image, Bard responded it would but then just created a new version of the image with sharper detail. Regardless, the feature is a nice addition.
Moving Up the Ranks
However, the most significant news about Bard this week may be its arrival at number two on the Chabots Arena leaderboard, hosted by the Large Model Systems Organization (LMSYS) run out of UC Berkeley. It trailed only OpenAI’s GPT-4 Turbo model. Chatbots Arena’s Elo score ranked Bard at 1215, with GPT-4 Turbo at 1249. These were the only models to exceed 1200 Elo. Gemini Pro (Dev API) and Gemini Pro Elo scores are 1122 and 1114, respectively.
You may wonder why Bard with Gemini Pro performs so much better than Gemini Pro (Dev API), ranked ninth, and Gemini Pro at twelfth. That is likely because of the type of benchmark hosted in Chatbots Arena. Since Bard is optimized for the chat use case, you should expect it to outperform an uncustomized foundation model.
What it Means
Gemini is Google’s LLM bet for the future. Its addition to Bard will likely improve the company's image among generative AI chat users in non-English languages. While you should not expect it to beat ChatGPT or Microsoft Copilot with GPT-4, the upgrade is sure to make it more competitive.
Bard does not appear to be widely adopted, and that was likely impacted by its later arrival in the market and limited promotion compared to its rivals. However, it was also due to less robust performance. Higher performance, as identified in the Chatbots Arena ranking, is a promising and critical development. The shift to Gemini Pro supporting 40 languages in Bard looks like a first step in a broader effort to drive global adoption. The chatbot wars are about to heat up.
