Welcoming Google Gemini Models to Airouter.io

In the ever-evolving landscape of language models, a few names stand out for their performance and promise. Google's Gemini models are certainly among them. I'm excited to share that the Gemini-1.5-Flash and Gemini-1.5-Pro models have officially joined the lineup at airouter.io.

These models are recognized for their excellence in the current rankings at LMSYS Arena, offering a compelling alternative to the widely-utilized OpenAI models. While the Pro version is honed for high-quality outputs, the Flash model strikes a balance between speed and cost-effectiveness while maintaining commendable quality. Early tests reveal that the Gemini Flash model performs exceptionally well when integrated into a model mix, a revelation that simply couldn't be overlooked.

Navigating the Integration Challenge

Integrating Google’s Gemini models into our algorithm wasn’t a straightforward task. The Gemini models come equipped with intriguing safety filters designed to filter out harassment, hate speech, sexually explicit, and dangerous content. On paper, this sounds perfect, but in practice, this feature has its quirks. Valid requests often get caught in this filter, even when we set the safety to "BLOCK_NONE". Occasionally, the model returns empty responses for safety reasons, unpredictably disrupting workflow.

Despite these hurdles, we've adapted. By enhancing our smart fallback strategies, these models can now be seamlessly used in production. When a blocking issue arises, our system gracefully defaults to the next best model, ensuring continuity and stability in our operations.

As a side note, we’re also gearing up to implement content moderation and injection attack detection capabilities on airouter.io.

Welcoming the Google Gemini models to airouter.io marks a step forward in improving the diversity and efficiency of our offerings, pushing the boundaries of what's achievable in AI-driven applications.