Google’s Gemini 3 Pro turns sparse MoE and 1M token context into a practical engine for multimodal agentic workloads

Google's Gemini 3 Pro is a significant step towards more general AI systems, with state-of-the-art reasoning, strong multimodal understanding, and improved agentic and vibe coding capabilities. The model is a sparse mixture of experts transformer with native multimodal support for text, images, audio, and video inputs. It can scale to 1M token contexts and generate up to 64k output tokens, making it suitable for code bases, long documents, or multi-hour transcripts.

Gemini 3 Pro is trained from scratch, covering large-scale public web text, code in many languages, images, audio, and video, combined with licensed data, user interaction data, and synthetic data. Post-training uses multimodal instruction tuning and reinforcement learning from human and critic feedback to improve multi-step reasoning, problem-solving, and theorem-proving behavior.

The system runs on Google Tensor Processing Units (TPUs), with training implemented in JAX and ML Pathways. On public benchmarks, Gemini 3 Pro clearly improves over Gemini 2.5 Pro and is competitive with other frontier models such as GPT 5.1 and Claude Sonnet 4.5.

Gemini 3 Pro achieves impressive results on various reasoning benchmarks and academic-style tasks, including:

• 37.5% score on Humanity's Last Exam without tools, compared to 21.6% for Gemini 2.5 Pro, 26.5% for GPT 5.1, and 13.7% for Claude Sonnet 4.5.
• 45.8% score on Humanity's Last Exam with search and code execution enabled.
• 31.1% score on ARC AGI 2 visual reasoning puzzles, up from 4.9% for Gemini 2.5 Pro, and ahead of GPT 5.1 at 17.6% and Claude Sonnet 4.5 at 13.6%.
• 91.9% score on GPQA Diamond, slightly ahead of GPT 5.1 at 88.1% and Claude Sonnet 4.5 at 83.4%.
• 95.0% score on AIME 2025 without tools and 100.0% with code execution.
• 23.4% score on MathArena Apex, a challenging contest-style benchmark.

Gemini 3 Pro is available in preview and is already integrated into various Google platforms, including the Gemini app, AI Mode in Search, Gemini API, Google AI Studio, Vertex AI, and the new Google Antigravity agentic development platform.