Google advances AI with Gemini 3.7 Flash and custom Frozen v2 chip

Google has released Gemini 3.7 Flash, its most advanced Flash model to date, specifically optimized for coding, UI design, and agentic workflows. The model demonstrates significant improvements in software engineering benchmarks, such as increasing its DeepSWE v1.1 score from 49% to 65.3%. Additionally, it has shown enhanced autonomous performance in business automation, scoring 30.4% on AutomationBench.

In parallel with software advancements, Alphabet is developing a new custom server processor codenamed ‘Frozen v2’. Designed to run Gemini models more efficiently, this hardware aims to integrate core model structural elements directly into the silicon. The project targets a 2028 deployment in data centers and promises to provide 6 to 10 times more data processing capacity per unit of energy compared to current advanced processors. The Frozen v2 chips are expected to work alongside Google’s existing Tensor Processing Unit (TPU) family rather than replacing it.