AB
AB
AiBoss

All Categories

project

Gemini 3 Deep Think - Google's dedicated inference model

Gemini 3 Deep Think is a dedicated inference model launched by Google DeepMind, focusing on deep scientific reasoning and complex problem solving. The model achieves 84.6% on the ARC-AGI-2 benchmark, far exceeding similar models.

What is Gemini 3 Deep Think?

Gemini 3 Deep Think is a dedicated inference model from Google DeepMind, focusing on deep scientific reasoning and complex problem solving. The model achieves 84.6% on the ARC-AGI-2 benchmark, far exceeding similar models; it scores 48.4% on Humanity’s Last Exam, and reaches an Elo score of 3455 on Codeforces, comparable to gold medal levels in physics and chemistry Olympiads. Designed specifically for research scenarios with "no clear rules and incomplete data," the model can discover logical flaws in mathematical papers, design semiconductor materials, and accelerate engineering prototype iterations. It is currently available to Google AI Ultra subscribers and early API applicants.

Key features of Gemini 3 Deep Think

  • Deep scientific reasoningThe model has reached the level of an Olympiad gold medalist in mathematics, physics, chemistry and other fields, and can discover hidden logical loopholes in academic papers.
  • Code generation and optimizationTransform complex ideas into high-quality executable code with the competitive programming skills of Elo 3455.
  • Engineering Prototype IterationThe model accelerates the mechanical hardware design cycle, enabling rapid physical prototype verification that is close to software development.
  • Multimodal research supportIt supports the integrated analysis of charts, literature, and experimental data, connecting interdisciplinary theories with practical applications.
  • 3D Design GenerationIt supports directly converting hand-drawn sketches into solid model files that can be used for 3D printing.

Gemini 3 Deep Think performance

  • Core benchmarks:
    • It achieved a score of 84.6% in the ARC-AGI-2 inference test, far exceeding Claude Opus 4.6's 68.8% and GPT-5.2's 52.9%.
    • It achieved a 48.4% success rate on the Humanity’s Last Exam problem set, significantly outperforming other cutting-edge models.
    • In Codeforces competitive programming rankings, Elo is ranked 3455, an improvement of nearly 1000 points compared to Gemini 3 Pro Preview.
  • Academic competitionsAchieve gold medal level in the written examinations of the 2025 International Mathematics, Physics, and Chemistry Olympiads. Receive a score of 50.5% in the Advanced Theoretical Physics CMT-Benchmark test.
  • Strong multimodal understandingMMMU-Pro achieved a multimodal benchmark score of 81.5%, slightly outperforming similar models.
  • Actual scientific research verificationIts value has been proven in real-world scenarios, successfully identifying logical flaws in mathematical papers that were missed by human peer reviewers.

How to use Gemini 3 Deep Think

  • Gemini App subscriptionGoogle AI Ultra subscribers can access Deep Think mode directly in the Gemini app without needing to apply for it separately.
  • API Early Access RequestResearchers, engineers, and companies can access the Gemini API through the early access program.Submit an application Obtain API call permissions.

Project address for Gemini 3 Deep Think

  • Project official website: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-deep-think/

Application scenarios of Gemini 3 Deep Think

  • Mathematical theoretical researchThe model can examine complex mathematical proofs and identify logical flaws that peer reviewers may have missed, thereby improving academic rigor.
  • Physical Chemistry SolutionIt supports the handling of highly challenging scientific problems at the level of the International Olympiad and assists in the accurate theoretical derivation.
  • Algorithm code developmentTransform complex ideas into efficient executable code to achieve top-level performance in competitive programming.
  • Mechanical Engineering IterationModels can accelerate hardware prototype design verification, enabling physical product iteration speeds to approach software development efficiency.
  • 3D printing manufacturingIt supports analyzing hand-drawn sketches and directly generating 3D printing model files that can be used in actual production.
Back