DeepMind Postpones Gemini 3.5 Pro Release, Discards Base Model

Google DeepMind has delayed the public launch of Gemini 3.5 Pro to mid-July. The primary factor for this postponement was the decision to scrap the base model, according to sources familiar with the matter. DeepMind subsequently initiated architectural changes for the model.

Originally, the new iteration was slated for public release in June. The delay directly impacts the model’s availability for external users and developers.

Internally, Google was developing both a base model and a refined, instruction-tuned version of Gemini 3.5 Pro. DeepMind ultimately decided against releasing the base model. The current focus is entirely on the instruction-tuned variant.

This decision prompted further pre-training and architectural adjustments. The instruction-tuned model, now the sole focus, is undergoing these additional development phases.

DeepMind's strategy for Gemini 3.5 Pro now centres on a singular, refined offering. This approach aims to deliver a more robust and capable model upon its eventual launch.

The revised timeline indicates a mid-July release. This provides DeepMind additional weeks to integrate the architectural changes and complete the necessary pre-training cycles. How will this focused development impact the model’s performance metrics?