Google has launched Gemini 3 Flash, an efficient AI tool designed for speed and affordability, matching the performance of its bigger sibling, Gemini 3 Pro. Marketed as a step towards democratizing frontier intelligence, it features a potentially groundbreaking architecture with a trillion parameters operating efficiently through extreme sparsity. This innovative approach decouples computational costs from total model capacity, activating only a fraction of parameters during inference.
Recent speculation links Gemini 3 Flash to a proposed 1.2 trillion-parameter model, suggesting it employs a learned routing technique called Parameter Efficient Expert Retrieval (PEER) for handling a vast number of experts efficiently. This model ranks third on the Artificial Analysis Intelligence Index, boasting the highest intelligence-per-dollar ratio despite a trade-off in token efficiency. Users note increased token usage and a tendency for “hallucinations,” with a 91% hallucination rate on uncertainties. Yet, its multimodal capabilities make it ideal for consumer applications, positioning it as a cost-effective choice for various tasks.
Source link
