双子座 2.5 深度思考并行化创造性问题解决

Gemini 2.5 Deep Think: A model within Google AI Ultra subscription designed for creative problem-solving using parallel thinking and extended inference time.
- Designed to better solve problems like iterative development, scientific discovery, and algorithm development that require creativity and strategic planning.
- An improved and optimized variation of the Deep Think model that achieved gold-medal performance at the 2025 International Mathematical Olympiad (IMO), solving five out of six problems perfectly and getting 35 points.
- The new variant trades some reasoning depth for speed, making it more practical for daily use compared to the previous version that took hours to reason about complex problems.
- Generates multiple alternative solutions in parallel and revises/combines them during reasoning to reach the best answer, requiring an extended reasoning time window and thus not suitable for interactive use like in a chat, and may have slowness or timeout issues.
- Uses novel reinforcement learning techniques during training to encourage using extended reasoning paths to test multiple hypotheses simultaneously and has access to a curated corpus of high-quality math problem solutions.
- Achieves state-of-the-art performance on several benchmarks including LiveCodeBench V6 and Humanity’s Last Exam.
Architecture and Features: Follows the sparse mixture-of-experts (MoE) architecture and has native multimodal support for text, vision, and audio inputs. Supports a 1M token input context window and 192K token output window.
Limitations: Early adopters pointed out a restricted number of queries available even to paying users, starting at five per day and later doubled to ten, which some took as a sign of the model's high running cost, possibly equivalent to running a large number of Gemini Pro models in parallel.
Release: In addition to the optimized Deep Think model for AI Ultra subscribers, Google is releasing the competition variant to the research community.