Conference Presentation, Keynote, Product Demonstration
What's New for Startups with Google DeepMind? | Omar Sanseviero | RAISE Summit 2026
- Google plans to release Gemini 2.5 Pro "coming soon" and expects to release multiple additional releases based on OVNI architecture within the next couple of months.
- The models are expected to reach a capability threshold where they can perform 95% of desired tasks, underpinned by a foundation that understands world physics and operational dynamics.
- Specific capabilities include grounding image generation via Google Search, initiating agentic tasks when appropriate, and performing reasoning on complex requests.
- OVNI models are defined as "anything-to-anything," with the initial iteration primarily supporting text, audio, and video-to-video inputs.
- The Live Translate API will enable real-time audio generation that occurs simultaneously with speech.
- Developer access is facilitated through ecosystem partners to minimize friction, while managed agents features aim to handle setup and sandboxing without user intervention.
- Gemma models are transitioning to a "proper Apache 2 license," a change expected to be of significant interest to many users.
- Hardware compatibility varies by model scale, with mid-sized models potentially running on normal gaming GPUs and the 31B model requiring high-end consumer GPUs.
- Context lengths that become "too large" will result in very large GPU requirements, necessitating careful selection between larger Pro models for complex, long-horizon tasks and other options.
- Users may utilize larger models like Pro for complex tasks or request granular control over specific image generation details.
- The development environment is described as the easiest it has ever been, with anticipation for the output generated by the community.