NEW Gemini 2.5 Pro Deep Think, Veo 3, Jules Coder, Gemma 3n, 2.5 Flash, & MORE!

Google I/O 2024: New Model Drops and Coding Agents

Key Concepts: Gemini 2.5 Pro Deep Think, Gemini 2.5 Flash, Gemma 3N, V3 (video generation model), Flow (text-to-film studio), Gemini Code Assist, Firebase Studio, Jules (coding agent), Google AI Ultra Plan.

Gemini 2.5 Pro Deep Think

• Main Point: A new version of Gemini 2.5 Pro, called Deep Think, significantly enhances reasoning capabilities.
• Details:
  • Simulates parallel hypothesis testing, allowing the model to "pause, think, and evaluate multiple pathways" before generating an answer.
  • Outpaces its predecessor (Gemini 2.5 Pro preview) in AI performance.
  • Achieves high scores on benchmarks: 84% on MMU for multimodal reasoning, excels at Live Codebench.
  • Features "thinking budgets" for controlled reasoning and "thought summaries" for transparency.
• Access: Available to trusted testers through the Gemini API, with broader access planned after further testing.
• Availability: Only accessible through the Google AI Ultra plan.

Google AI Ultra Plan

• Main Point: A new subscription plan is required to access the Gemini 2.5 Pro Deep Think model.
• Details:
  • Cost: $249.99 per month (first 3 months at $124.99).
  • Includes access to Gemini 2.5 Pro Deep Think, V3 model, and Flow.
  • Currently only available in the US, with more countries coming soon.
  • The Google AI Pro plan ($20/month) does not grant access to Deep Think.

Gemini 2.5 Flash

• Main Point: A faster, smarter, and cheaper model optimized for low latency and cost efficiency.
• Details:
  • A "lean, high-speed sibling" of Gemini 2.5 Pro.
  • Uses 20-30% fewer tokens for the same tasks.
  • Supports long context, multimodal input, and reasoning tasks.
  • Features native audio output and multi-speaker text-to-speech integration.
  • Includes boosted security against prompt injection.
• Performance: Competitive performance in reasoning and science, slightly behind in coding compared to other state-of-the-art models (OpenAI's GPT-4, Claude 3.5 Sonnet, Grok-1.5, DeepSeek R1).
• Availability: Available in Google AI Studio, the Gemini app, and soon through Vertex AI.

Gemma 3N

• Main Point: A tiny, multimodal model for mobile and edge devices.
• Details:
  • A 4 billion parameter model.
  • Supports text, image, audio, and video.
  • Optimized for smartphones and edge devices.
  • Performance is on par with Claude 3.5 Sonnet.
  • Outperforms GPT-4.0 Nano, Llama 3 Maverick, and Phi-3-mini, despite being smaller.
• Applications: Ideal for on-device AI tasks like AR overlays, instant translations, and personal assistants.

V3 (Video Generation Model) and Flow

• Main Point: V3 is a high-fidelity video generation model with sound and dialogue, coupled with Flow, a text-to-film studio.
• Details:
  • V3 generates 4K realism videos with native sound, dialogue, and ambient noise.
  • Designed for storytellers, educators, marketers, and content creators.
  • Can be paired with Gemini to generate videos from structured prompts.
  • Example: The video showcases a scene with realistic dialogue.
  • Flow is a new creative tool that combines V3 with Gemini to automate film scene creation from text prompts.

Gemini Code Assist

• Main Point: A free AI coding companion with enhanced capabilities through the 2.5 upgrade.
• Details:
  • Supports the new Gemini 2.5 Pro and will support Deep Think when fully available.
  • Offers a 2 million token context for larger codebases.
  • Provides code reviews, inline suggestions, and debugging tips.
  • Automatically detects and repairs bugs inside Google Colab Notebooks.

Firebase Studio

• Main Point: Converts Figma designs into functional frontends in minutes.
• Details:
  • Autogenerates backends, UI systems, and databases from Figma designs.
  • Uses Gemini 2.5 Pro to optimize layout and logic.

Jules (Coding Agent)

• Main Point: A new coding agent that autonomously handles bug fixes, refactors, and feature prototyping.
• Details:
  • A competitor to OpenAI's CodeX.
  • Works asynchronously with the codebase.
  • Operates off of Gemini 2.5 capabilities with tool use.
  • Users write the problem, and Jules creates a solution and submits PRs.
  • Functions as an AI developer that takes control of tasks autonomously.

Additional Updates (Mentioned in Twitter Thread)

• New diffusion model (ImageGen 4) to rival OpenAI's image generation model.

Conclusion

Google I/O 2024 showcased significant advancements in AI models and tools, particularly with the introduction of Gemini 2.5 Pro Deep Think (accessible via the new Ultra subscription), Gemini 2.5 Flash, Gemma 3N, and the V3 video generation model. The updates to Gemini Code Assist, Firebase Studio, and the introduction of Jules further demonstrate Google's commitment to enhancing developer workflows and AI-powered creative tools. While some features are initially limited to the US and require a premium subscription, the overall announcements highlight Google's continued innovation in the AI space.